blob: 257b61789eeba9fd064470c91e3a8bffb8d1812d [file] [log] [blame]
Linus Torvalds1da177e2005-04-16 15:20:36 -07001/*
2 * INET An implementation of the TCP/IP protocol suite for the LINUX
3 * operating system. INET is implemented using the BSD Socket
4 * interface as the means of communication with the user level.
5 *
6 * Implementation of the Transmission Control Protocol(TCP).
7 *
Jesper Juhl02c30a82005-05-05 16:16:16 -07008 * Authors: Ross Biro
Linus Torvalds1da177e2005-04-16 15:20:36 -07009 * Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
10 * Mark Evans, <evansmp@uhura.aston.ac.uk>
11 * Corey Minyard <wf-rch!minyard@relay.EU.net>
12 * Florian La Roche, <flla@stud.uni-sb.de>
13 * Charles Hedrick, <hedrick@klinzhai.rutgers.edu>
14 * Linus Torvalds, <torvalds@cs.helsinki.fi>
15 * Alan Cox, <gw4pts@gw4pts.ampr.org>
16 * Matthew Dillon, <dillon@apollo.west.oic.com>
17 * Arnt Gulbrandsen, <agulbra@nvg.unit.no>
18 * Jorge Cwik, <jorge@laser.satlink.net>
19 */
20
21/*
22 * Changes:
23 * Pedro Roque : Fast Retransmit/Recovery.
24 * Two receive queues.
25 * Retransmit queue handled by TCP.
26 * Better retransmit timer handling.
27 * New congestion avoidance.
28 * Header prediction.
29 * Variable renaming.
30 *
31 * Eric : Fast Retransmit.
32 * Randy Scott : MSS option defines.
33 * Eric Schenk : Fixes to slow start algorithm.
34 * Eric Schenk : Yet another double ACK bug.
35 * Eric Schenk : Delayed ACK bug fixes.
36 * Eric Schenk : Floyd style fast retrans war avoidance.
37 * David S. Miller : Don't allow zero congestion window.
38 * Eric Schenk : Fix retransmitter so that it sends
39 * next packet on ack of previous packet.
40 * Andi Kleen : Moved open_request checking here
41 * and process RSTs for open_requests.
42 * Andi Kleen : Better prune_queue, and other fixes.
Stephen Hemmingercaa20d9a2005-11-10 17:13:47 -080043 * Andrey Savochkin: Fix RTT measurements in the presence of
Linus Torvalds1da177e2005-04-16 15:20:36 -070044 * timestamps.
45 * Andrey Savochkin: Check sequence numbers correctly when
46 * removing SACKs due to in sequence incoming
47 * data segments.
48 * Andi Kleen: Make sure we never ack data there is not
49 * enough room for. Also make this condition
50 * a fatal error if it might still happen.
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +090051 * Andi Kleen: Add tcp_measure_rcv_mss to make
Linus Torvalds1da177e2005-04-16 15:20:36 -070052 * connections with MSS<min(MTU,ann. MSS)
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +090053 * work without delayed acks.
Linus Torvalds1da177e2005-04-16 15:20:36 -070054 * Andi Kleen: Process packets with PSH set in the
55 * fast path.
56 * J Hadi Salim: ECN support
57 * Andrei Gurtov,
58 * Pasi Sarolahti,
59 * Panu Kuhlberg: Experimental audit of TCP (re)transmission
60 * engine. Lots of bugs are found.
61 * Pasi Sarolahti: F-RTO for dealing with spurious RTOs
Linus Torvalds1da177e2005-04-16 15:20:36 -070062 */
63
Joe Perchesafd465032012-03-12 07:03:32 +000064#define pr_fmt(fmt) "TCP: " fmt
65
Linus Torvalds1da177e2005-04-16 15:20:36 -070066#include <linux/mm.h>
Tejun Heo5a0e3ad2010-03-24 17:04:11 +090067#include <linux/slab.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070068#include <linux/module.h>
69#include <linux/sysctl.h>
Ilpo Järvinena0bffff2009-03-21 13:36:17 -070070#include <linux/kernel.h>
Satoru SATOH5ffc02a2008-05-04 22:14:42 -070071#include <net/dst.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070072#include <net/tcp.h>
73#include <net/inet_common.h>
74#include <linux/ipsec.h>
75#include <asm/unaligned.h>
Chris Leech1a2449a2006-05-23 18:05:53 -070076#include <net/netdma.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070077
Brian Haleyab32ea52006-09-22 14:15:41 -070078int sysctl_tcp_timestamps __read_mostly = 1;
79int sysctl_tcp_window_scaling __read_mostly = 1;
80int sysctl_tcp_sack __read_mostly = 1;
81int sysctl_tcp_fack __read_mostly = 1;
82int sysctl_tcp_reordering __read_mostly = TCP_FASTRETRANS_THRESH;
Eric Dumazet4bc2f182010-07-09 21:22:10 +000083EXPORT_SYMBOL(sysctl_tcp_reordering);
Ilpo Järvinen255cac92009-05-04 11:07:36 -070084int sysctl_tcp_ecn __read_mostly = 2;
Eric Dumazet4bc2f182010-07-09 21:22:10 +000085EXPORT_SYMBOL(sysctl_tcp_ecn);
Brian Haleyab32ea52006-09-22 14:15:41 -070086int sysctl_tcp_dsack __read_mostly = 1;
87int sysctl_tcp_app_win __read_mostly = 31;
Eric Dumazetb49960a2012-05-02 02:28:41 +000088int sysctl_tcp_adv_win_scale __read_mostly = 1;
Eric Dumazet4bc2f182010-07-09 21:22:10 +000089EXPORT_SYMBOL(sysctl_tcp_adv_win_scale);
Linus Torvalds1da177e2005-04-16 15:20:36 -070090
Brian Haleyab32ea52006-09-22 14:15:41 -070091int sysctl_tcp_stdurg __read_mostly;
92int sysctl_tcp_rfc1337 __read_mostly;
93int sysctl_tcp_max_orphans __read_mostly = NR_FILE;
Ilpo Järvinenc96fd3d2007-09-20 11:36:37 -070094int sysctl_tcp_frto __read_mostly = 2;
Ilpo Järvinen3cfe3ba2007-02-27 10:09:49 -080095int sysctl_tcp_frto_response __read_mostly;
Brian Haleyab32ea52006-09-22 14:15:41 -070096int sysctl_tcp_nometrics_save __read_mostly;
Linus Torvalds1da177e2005-04-16 15:20:36 -070097
Andreas Petlund7e380172010-02-18 04:48:19 +000098int sysctl_tcp_thin_dupack __read_mostly;
99
Brian Haleyab32ea52006-09-22 14:15:41 -0700100int sysctl_tcp_moderate_rcvbuf __read_mostly = 1;
101int sysctl_tcp_abc __read_mostly;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700102
Linus Torvalds1da177e2005-04-16 15:20:36 -0700103#define FLAG_DATA 0x01 /* Incoming frame contained data. */
104#define FLAG_WIN_UPDATE 0x02 /* Incoming ACK was a window update. */
105#define FLAG_DATA_ACKED 0x04 /* This ACK acknowledged new data. */
106#define FLAG_RETRANS_DATA_ACKED 0x08 /* "" "" some of which was retransmitted. */
107#define FLAG_SYN_ACKED 0x10 /* This ACK acknowledged SYN. */
108#define FLAG_DATA_SACKED 0x20 /* New SACK. */
109#define FLAG_ECE 0x40 /* ECE in this ACK */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700110#define FLAG_SLOWPATH 0x100 /* Do not skip RFC checks for window update.*/
Ilpo Järvinen4dc26652007-02-21 23:16:11 -0800111#define FLAG_ONLY_ORIG_SACKED 0x200 /* SACKs only non-rexmit sent before RTO */
Ilpo Järvinen2e605292007-08-02 19:46:58 -0700112#define FLAG_SND_UNA_ADVANCED 0x400 /* Snd_una was changed (!= FLAG_DATA_ACKED) */
Ryousei Takano564262c2007-10-25 23:03:52 -0700113#define FLAG_DSACKING_ACK 0x800 /* SACK blocks contained D-SACK info */
Ilpo Järvinen009a2e32007-09-20 11:34:38 -0700114#define FLAG_NONHEAD_RETRANS_ACKED 0x1000 /* Non-head rexmitted data was ACKed */
Ilpo Järvinencadbd032007-12-31 04:49:21 -0800115#define FLAG_SACK_RENEGING 0x2000 /* snd_una advanced to a sacked seq */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700116
117#define FLAG_ACKED (FLAG_DATA_ACKED|FLAG_SYN_ACKED)
118#define FLAG_NOT_DUP (FLAG_DATA|FLAG_WIN_UPDATE|FLAG_ACKED)
119#define FLAG_CA_ALERT (FLAG_DATA_SACKED|FLAG_ECE)
120#define FLAG_FORWARD_PROGRESS (FLAG_ACKED|FLAG_DATA_SACKED)
Ilpo Järvinen2e605292007-08-02 19:46:58 -0700121#define FLAG_ANY_PROGRESS (FLAG_FORWARD_PROGRESS|FLAG_SND_UNA_ADVANCED)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700122
Linus Torvalds1da177e2005-04-16 15:20:36 -0700123#define TCP_REMNANT (TCP_FLAG_FIN|TCP_FLAG_URG|TCP_FLAG_SYN|TCP_FLAG_PSH)
Ilpo Järvinenbdf1ee52007-05-27 02:04:16 -0700124#define TCP_HP_BITS (~(TCP_RESERVED_BITS|TCP_FLAG_PSH))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700125
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +0900126/* Adapt the MSS value used to make delayed ack decision to the
Linus Torvalds1da177e2005-04-16 15:20:36 -0700127 * real world.
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +0900128 */
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800129static void tcp_measure_rcv_mss(struct sock *sk, const struct sk_buff *skb)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700130{
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700131 struct inet_connection_sock *icsk = inet_csk(sk);
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +0900132 const unsigned int lss = icsk->icsk_ack.last_seg_size;
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700133 unsigned int len;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700134
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +0900135 icsk->icsk_ack.last_seg_size = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700136
137 /* skb->len may jitter because of SACKs, even if peer
138 * sends good full-sized frames.
139 */
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800140 len = skb_shinfo(skb)->gso_size ? : skb->len;
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700141 if (len >= icsk->icsk_ack.rcv_mss) {
142 icsk->icsk_ack.rcv_mss = len;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700143 } else {
144 /* Otherwise, we make more careful check taking into account,
145 * that SACKs block is variable.
146 *
147 * "len" is invariant segment length, including TCP header.
148 */
Arnaldo Carvalho de Melo9c702202007-04-25 18:04:18 -0700149 len += skb->data - skb_transport_header(skb);
William Allen Simpsonbee7ca92009-11-10 09:51:18 +0000150 if (len >= TCP_MSS_DEFAULT + sizeof(struct tcphdr) ||
Linus Torvalds1da177e2005-04-16 15:20:36 -0700151 /* If PSH is not set, packet should be
152 * full sized, provided peer TCP is not badly broken.
153 * This observation (if it is correct 8)) allows
154 * to handle super-low mtu links fairly.
155 */
156 (len >= TCP_MIN_MSS + sizeof(struct tcphdr) &&
Arnaldo Carvalho de Meloaa8223c2007-04-10 21:04:22 -0700157 !(tcp_flag_word(tcp_hdr(skb)) & TCP_REMNANT))) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700158 /* Subtract also invariant (if peer is RFC compliant),
159 * tcp header plus fixed timestamp option length.
160 * Resulting "len" is MSS free of SACK jitter.
161 */
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700162 len -= tcp_sk(sk)->tcp_header_len;
163 icsk->icsk_ack.last_seg_size = len;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700164 if (len == lss) {
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700165 icsk->icsk_ack.rcv_mss = len;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700166 return;
167 }
168 }
Alexey Kuznetsov1ef96962006-09-19 12:52:50 -0700169 if (icsk->icsk_ack.pending & ICSK_ACK_PUSHED)
170 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED2;
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700171 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700172 }
173}
174
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700175static void tcp_incr_quickack(struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700176{
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700177 struct inet_connection_sock *icsk = inet_csk(sk);
178 unsigned quickacks = tcp_sk(sk)->rcv_wnd / (2 * icsk->icsk_ack.rcv_mss);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700179
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800180 if (quickacks == 0)
181 quickacks = 2;
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700182 if (quickacks > icsk->icsk_ack.quick)
183 icsk->icsk_ack.quick = min(quickacks, TCP_MAX_QUICKACKS);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700184}
185
stephen hemminger1b9f4092010-09-28 19:30:14 +0000186static void tcp_enter_quickack_mode(struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700187{
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700188 struct inet_connection_sock *icsk = inet_csk(sk);
189 tcp_incr_quickack(sk);
190 icsk->icsk_ack.pingpong = 0;
191 icsk->icsk_ack.ato = TCP_ATO_MIN;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700192}
193
194/* Send ACKs quickly, if "quick" count is not exhausted
195 * and the session is not interactive.
196 */
197
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700198static inline int tcp_in_quickack_mode(const struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700199{
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700200 const struct inet_connection_sock *icsk = inet_csk(sk);
201 return icsk->icsk_ack.quick && !icsk->icsk_ack.pingpong;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700202}
203
Ilpo Järvinenbdf1ee52007-05-27 02:04:16 -0700204static inline void TCP_ECN_queue_cwr(struct tcp_sock *tp)
205{
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800206 if (tp->ecn_flags & TCP_ECN_OK)
Ilpo Järvinenbdf1ee52007-05-27 02:04:16 -0700207 tp->ecn_flags |= TCP_ECN_QUEUE_CWR;
208}
209
Eric Dumazetcf533ea2011-10-21 05:22:42 -0400210static inline void TCP_ECN_accept_cwr(struct tcp_sock *tp, const struct sk_buff *skb)
Ilpo Järvinenbdf1ee52007-05-27 02:04:16 -0700211{
212 if (tcp_hdr(skb)->cwr)
213 tp->ecn_flags &= ~TCP_ECN_DEMAND_CWR;
214}
215
216static inline void TCP_ECN_withdraw_cwr(struct tcp_sock *tp)
217{
218 tp->ecn_flags &= ~TCP_ECN_DEMAND_CWR;
219}
220
Eric Dumazet7a269ff2011-09-22 20:02:19 +0000221static inline void TCP_ECN_check_ce(struct tcp_sock *tp, const struct sk_buff *skb)
Ilpo Järvinenbdf1ee52007-05-27 02:04:16 -0700222{
Eric Dumazet7a269ff2011-09-22 20:02:19 +0000223 if (!(tp->ecn_flags & TCP_ECN_OK))
224 return;
225
Eric Dumazetb82d1bb2011-09-27 02:20:08 -0400226 switch (TCP_SKB_CB(skb)->ip_dsfield & INET_ECN_MASK) {
Eric Dumazet7a269ff2011-09-22 20:02:19 +0000227 case INET_ECN_NOT_ECT:
Ilpo Järvinenbdf1ee52007-05-27 02:04:16 -0700228 /* Funny extension: if ECT is not set on a segment,
Eric Dumazet7a269ff2011-09-22 20:02:19 +0000229 * and we already seen ECT on a previous segment,
230 * it is probably a retransmit.
231 */
232 if (tp->ecn_flags & TCP_ECN_SEEN)
Ilpo Järvinenbdf1ee52007-05-27 02:04:16 -0700233 tcp_enter_quickack_mode((struct sock *)tp);
Eric Dumazet7a269ff2011-09-22 20:02:19 +0000234 break;
235 case INET_ECN_CE:
236 tp->ecn_flags |= TCP_ECN_DEMAND_CWR;
237 /* fallinto */
238 default:
239 tp->ecn_flags |= TCP_ECN_SEEN;
Ilpo Järvinenbdf1ee52007-05-27 02:04:16 -0700240 }
241}
242
Eric Dumazetcf533ea2011-10-21 05:22:42 -0400243static inline void TCP_ECN_rcv_synack(struct tcp_sock *tp, const struct tcphdr *th)
Ilpo Järvinenbdf1ee52007-05-27 02:04:16 -0700244{
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800245 if ((tp->ecn_flags & TCP_ECN_OK) && (!th->ece || th->cwr))
Ilpo Järvinenbdf1ee52007-05-27 02:04:16 -0700246 tp->ecn_flags &= ~TCP_ECN_OK;
247}
248
Eric Dumazetcf533ea2011-10-21 05:22:42 -0400249static inline void TCP_ECN_rcv_syn(struct tcp_sock *tp, const struct tcphdr *th)
Ilpo Järvinenbdf1ee52007-05-27 02:04:16 -0700250{
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800251 if ((tp->ecn_flags & TCP_ECN_OK) && (!th->ece || !th->cwr))
Ilpo Järvinenbdf1ee52007-05-27 02:04:16 -0700252 tp->ecn_flags &= ~TCP_ECN_OK;
253}
254
Eric Dumazetcf533ea2011-10-21 05:22:42 -0400255static inline int TCP_ECN_rcv_ecn_echo(const struct tcp_sock *tp, const struct tcphdr *th)
Ilpo Järvinenbdf1ee52007-05-27 02:04:16 -0700256{
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800257 if (th->ece && !th->syn && (tp->ecn_flags & TCP_ECN_OK))
Ilpo Järvinenbdf1ee52007-05-27 02:04:16 -0700258 return 1;
259 return 0;
260}
261
Linus Torvalds1da177e2005-04-16 15:20:36 -0700262/* Buffer size and advertised window tuning.
263 *
264 * 1. Tuning sk->sk_sndbuf, when connection enters established state.
265 */
266
267static void tcp_fixup_sndbuf(struct sock *sk)
268{
Eric Dumazet87fb4b72011-10-13 07:28:54 +0000269 int sndmem = SKB_TRUESIZE(tcp_sk(sk)->rx_opt.mss_clamp + MAX_TCP_HEADER);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700270
Eric Dumazet06a59ec2011-10-13 18:24:42 +0000271 sndmem *= TCP_INIT_CWND;
272 if (sk->sk_sndbuf < sndmem)
273 sk->sk_sndbuf = min(sndmem, sysctl_tcp_wmem[2]);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700274}
275
276/* 2. Tuning advertised window (window_clamp, rcv_ssthresh)
277 *
278 * All tcp_full_space() is split to two parts: "network" buffer, allocated
279 * forward and advertised in receiver window (tp->rcv_wnd) and
280 * "application buffer", required to isolate scheduling/application
281 * latencies from network.
282 * window_clamp is maximal advertised window. It can be less than
283 * tcp_full_space(), in this case tcp_full_space() - window_clamp
284 * is reserved for "application" buffer. The less window_clamp is
285 * the smoother our behaviour from viewpoint of network, but the lower
286 * throughput and the higher sensitivity of the connection to losses. 8)
287 *
288 * rcv_ssthresh is more strict window_clamp used at "slow start"
289 * phase to predict further behaviour of this connection.
290 * It is used for two goals:
291 * - to enforce header prediction at sender, even when application
292 * requires some significant "application buffer". It is check #1.
293 * - to prevent pruning of receive queue because of misprediction
294 * of receiver window. Check #2.
295 *
296 * The scheme does not work when sender sends good segments opening
Stephen Hemmingercaa20d9a2005-11-10 17:13:47 -0800297 * window and then starts to feed us spaghetti. But it should work
Linus Torvalds1da177e2005-04-16 15:20:36 -0700298 * in common situations. Otherwise, we have to rely on queue collapsing.
299 */
300
301/* Slow part of check#2. */
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -0700302static int __tcp_grow_window(const struct sock *sk, const struct sk_buff *skb)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700303{
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -0700304 struct tcp_sock *tp = tcp_sk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700305 /* Optimize this! */
Eric Dumazetdfd4f0a2007-12-21 06:07:53 -0800306 int truesize = tcp_win_from_space(skb->truesize) >> 1;
307 int window = tcp_win_from_space(sysctl_tcp_rmem[2]) >> 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700308
309 while (tp->rcv_ssthresh <= window) {
310 if (truesize <= skb->len)
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700311 return 2 * inet_csk(sk)->icsk_ack.rcv_mss;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700312
313 truesize >>= 1;
314 window >>= 1;
315 }
316 return 0;
317}
318
Eric Dumazetcf533ea2011-10-21 05:22:42 -0400319static void tcp_grow_window(struct sock *sk, const struct sk_buff *skb)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700320{
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -0700321 struct tcp_sock *tp = tcp_sk(sk);
322
Linus Torvalds1da177e2005-04-16 15:20:36 -0700323 /* Check #1 */
324 if (tp->rcv_ssthresh < tp->window_clamp &&
325 (int)tp->rcv_ssthresh < tcp_space(sk) &&
Glauber Costa180d8cd2011-12-11 21:47:02 +0000326 !sk_under_memory_pressure(sk)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700327 int incr;
328
329 /* Check #2. Increase window, if skb with such overhead
330 * will fit to rcvbuf in future.
331 */
332 if (tcp_win_from_space(skb->truesize) <= skb->len)
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800333 incr = 2 * tp->advmss;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700334 else
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -0700335 incr = __tcp_grow_window(sk, skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700336
337 if (incr) {
Eric Dumazet4d846f02012-04-16 23:28:07 +0000338 incr = max_t(int, incr, 2 * skb->len);
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800339 tp->rcv_ssthresh = min(tp->rcv_ssthresh + incr,
340 tp->window_clamp);
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700341 inet_csk(sk)->icsk_ack.quick |= 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700342 }
343 }
344}
345
346/* 3. Tuning rcvbuf, when connection enters established state. */
347
348static void tcp_fixup_rcvbuf(struct sock *sk)
349{
Eric Dumazete9266a02011-10-20 16:53:56 -0400350 u32 mss = tcp_sk(sk)->advmss;
351 u32 icwnd = TCP_DEFAULT_INIT_RCVWND;
352 int rcvmem;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700353
Eric Dumazete9266a02011-10-20 16:53:56 -0400354 /* Limit to 10 segments if mss <= 1460,
355 * or 14600/mss segments, with a minimum of two segments.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700356 */
Eric Dumazete9266a02011-10-20 16:53:56 -0400357 if (mss > 1460)
358 icwnd = max_t(u32, (1460 * TCP_DEFAULT_INIT_RCVWND) / mss, 2);
359
360 rcvmem = SKB_TRUESIZE(mss + MAX_TCP_HEADER);
361 while (tcp_win_from_space(rcvmem) < mss)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700362 rcvmem += 128;
Eric Dumazete9266a02011-10-20 16:53:56 -0400363
364 rcvmem *= icwnd;
365
366 if (sk->sk_rcvbuf < rcvmem)
367 sk->sk_rcvbuf = min(rcvmem, sysctl_tcp_rmem[2]);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700368}
369
Stephen Hemmingercaa20d9a2005-11-10 17:13:47 -0800370/* 4. Try to fixup all. It is made immediately after connection enters
Linus Torvalds1da177e2005-04-16 15:20:36 -0700371 * established state.
372 */
373static void tcp_init_buffer_space(struct sock *sk)
374{
375 struct tcp_sock *tp = tcp_sk(sk);
376 int maxwin;
377
378 if (!(sk->sk_userlocks & SOCK_RCVBUF_LOCK))
379 tcp_fixup_rcvbuf(sk);
380 if (!(sk->sk_userlocks & SOCK_SNDBUF_LOCK))
381 tcp_fixup_sndbuf(sk);
382
383 tp->rcvq_space.space = tp->rcv_wnd;
384
385 maxwin = tcp_full_space(sk);
386
387 if (tp->window_clamp >= maxwin) {
388 tp->window_clamp = maxwin;
389
390 if (sysctl_tcp_app_win && maxwin > 4 * tp->advmss)
391 tp->window_clamp = max(maxwin -
392 (maxwin >> sysctl_tcp_app_win),
393 4 * tp->advmss);
394 }
395
396 /* Force reservation of one segment. */
397 if (sysctl_tcp_app_win &&
398 tp->window_clamp > 2 * tp->advmss &&
399 tp->window_clamp + tp->advmss > maxwin)
400 tp->window_clamp = max(2 * tp->advmss, maxwin - tp->advmss);
401
402 tp->rcv_ssthresh = min(tp->rcv_ssthresh, tp->window_clamp);
403 tp->snd_cwnd_stamp = tcp_time_stamp;
404}
405
Linus Torvalds1da177e2005-04-16 15:20:36 -0700406/* 5. Recalculate window clamp after socket hit its memory bounds. */
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -0700407static void tcp_clamp_window(struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700408{
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -0700409 struct tcp_sock *tp = tcp_sk(sk);
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -0300410 struct inet_connection_sock *icsk = inet_csk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700411
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -0300412 icsk->icsk_ack.quick = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700413
John Heffner326f36e2005-11-10 17:11:48 -0800414 if (sk->sk_rcvbuf < sysctl_tcp_rmem[2] &&
415 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK) &&
Glauber Costa180d8cd2011-12-11 21:47:02 +0000416 !sk_under_memory_pressure(sk) &&
417 sk_memory_allocated(sk) < sk_prot_mem_limits(sk, 0)) {
John Heffner326f36e2005-11-10 17:11:48 -0800418 sk->sk_rcvbuf = min(atomic_read(&sk->sk_rmem_alloc),
419 sysctl_tcp_rmem[2]);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700420 }
John Heffner326f36e2005-11-10 17:11:48 -0800421 if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf)
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800422 tp->rcv_ssthresh = min(tp->window_clamp, 2U * tp->advmss);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700423}
424
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800425/* Initialize RCV_MSS value.
426 * RCV_MSS is an our guess about MSS used by the peer.
427 * We haven't any direct information about the MSS.
428 * It's better to underestimate the RCV_MSS rather than overestimate.
429 * Overestimations make us ACKing less frequently than needed.
430 * Underestimations are more easy to detect and fix by tcp_measure_rcv_mss().
431 */
432void tcp_initialize_rcv_mss(struct sock *sk)
433{
Eric Dumazetcf533ea2011-10-21 05:22:42 -0400434 const struct tcp_sock *tp = tcp_sk(sk);
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800435 unsigned int hint = min_t(unsigned int, tp->advmss, tp->mss_cache);
436
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800437 hint = min(hint, tp->rcv_wnd / 2);
William Allen Simpsonbee7ca92009-11-10 09:51:18 +0000438 hint = min(hint, TCP_MSS_DEFAULT);
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800439 hint = max(hint, TCP_MIN_MSS);
440
441 inet_csk(sk)->icsk_ack.rcv_mss = hint;
442}
Eric Dumazet4bc2f182010-07-09 21:22:10 +0000443EXPORT_SYMBOL(tcp_initialize_rcv_mss);
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800444
Linus Torvalds1da177e2005-04-16 15:20:36 -0700445/* Receiver "autotuning" code.
446 *
447 * The algorithm for RTT estimation w/o timestamps is based on
448 * Dynamic Right-Sizing (DRS) by Wu Feng and Mike Fisk of LANL.
Justin P. Mattock631dd1a2010-10-18 11:03:14 +0200449 * <http://public.lanl.gov/radiant/pubs.html#DRS>
Linus Torvalds1da177e2005-04-16 15:20:36 -0700450 *
451 * More detail on this code can be found at
Justin P. Mattock631dd1a2010-10-18 11:03:14 +0200452 * <http://staff.psc.edu/jheffner/>,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700453 * though this reference is out of date. A new paper
454 * is pending.
455 */
456static void tcp_rcv_rtt_update(struct tcp_sock *tp, u32 sample, int win_dep)
457{
458 u32 new_sample = tp->rcv_rtt_est.rtt;
459 long m = sample;
460
461 if (m == 0)
462 m = 1;
463
464 if (new_sample != 0) {
465 /* If we sample in larger samples in the non-timestamp
466 * case, we could grossly overestimate the RTT especially
467 * with chatty applications or bulk transfer apps which
468 * are stalled on filesystem I/O.
469 *
470 * Also, since we are only going for a minimum in the
Stephen Hemminger31f34262005-11-15 15:17:10 -0800471 * non-timestamp case, we do not smooth things out
Stephen Hemmingercaa20d9a2005-11-10 17:13:47 -0800472 * else with timestamps disabled convergence takes too
Linus Torvalds1da177e2005-04-16 15:20:36 -0700473 * long.
474 */
475 if (!win_dep) {
476 m -= (new_sample >> 3);
477 new_sample += m;
Neal Cardwell18a223e2012-04-10 07:59:20 +0000478 } else {
479 m <<= 3;
480 if (m < new_sample)
481 new_sample = m;
482 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700483 } else {
Stephen Hemmingercaa20d9a2005-11-10 17:13:47 -0800484 /* No previous measure. */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700485 new_sample = m << 3;
486 }
487
488 if (tp->rcv_rtt_est.rtt != new_sample)
489 tp->rcv_rtt_est.rtt = new_sample;
490}
491
492static inline void tcp_rcv_rtt_measure(struct tcp_sock *tp)
493{
494 if (tp->rcv_rtt_est.time == 0)
495 goto new_measure;
496 if (before(tp->rcv_nxt, tp->rcv_rtt_est.seq))
497 return;
Neal Cardwell651913c2012-04-27 11:29:37 -0400498 tcp_rcv_rtt_update(tp, tcp_time_stamp - tp->rcv_rtt_est.time, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700499
500new_measure:
501 tp->rcv_rtt_est.seq = tp->rcv_nxt + tp->rcv_wnd;
502 tp->rcv_rtt_est.time = tcp_time_stamp;
503}
504
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800505static inline void tcp_rcv_rtt_measure_ts(struct sock *sk,
506 const struct sk_buff *skb)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700507{
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700508 struct tcp_sock *tp = tcp_sk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700509 if (tp->rx_opt.rcv_tsecr &&
510 (TCP_SKB_CB(skb)->end_seq -
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700511 TCP_SKB_CB(skb)->seq >= inet_csk(sk)->icsk_ack.rcv_mss))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700512 tcp_rcv_rtt_update(tp, tcp_time_stamp - tp->rx_opt.rcv_tsecr, 0);
513}
514
515/*
516 * This function should be called every time data is copied to user space.
517 * It calculates the appropriate TCP receive buffer space.
518 */
519void tcp_rcv_space_adjust(struct sock *sk)
520{
521 struct tcp_sock *tp = tcp_sk(sk);
522 int time;
523 int space;
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +0900524
Linus Torvalds1da177e2005-04-16 15:20:36 -0700525 if (tp->rcvq_space.time == 0)
526 goto new_measure;
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +0900527
Linus Torvalds1da177e2005-04-16 15:20:36 -0700528 time = tcp_time_stamp - tp->rcvq_space.time;
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800529 if (time < (tp->rcv_rtt_est.rtt >> 3) || tp->rcv_rtt_est.rtt == 0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700530 return;
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +0900531
Linus Torvalds1da177e2005-04-16 15:20:36 -0700532 space = 2 * (tp->copied_seq - tp->rcvq_space.seq);
533
534 space = max(tp->rcvq_space.space, space);
535
536 if (tp->rcvq_space.space != space) {
537 int rcvmem;
538
539 tp->rcvq_space.space = space;
540
John Heffner6fcf9412006-02-09 17:06:57 -0800541 if (sysctl_tcp_moderate_rcvbuf &&
542 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700543 int new_clamp = space;
544
545 /* Receive space grows, normalize in order to
546 * take into account packet headers and sk_buff
547 * structure overhead.
548 */
549 space /= tp->advmss;
550 if (!space)
551 space = 1;
Eric Dumazet87fb4b72011-10-13 07:28:54 +0000552 rcvmem = SKB_TRUESIZE(tp->advmss + MAX_TCP_HEADER);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700553 while (tcp_win_from_space(rcvmem) < tp->advmss)
554 rcvmem += 128;
555 space *= rcvmem;
556 space = min(space, sysctl_tcp_rmem[2]);
557 if (space > sk->sk_rcvbuf) {
558 sk->sk_rcvbuf = space;
559
560 /* Make the window clamp follow along. */
561 tp->window_clamp = new_clamp;
562 }
563 }
564 }
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +0900565
Linus Torvalds1da177e2005-04-16 15:20:36 -0700566new_measure:
567 tp->rcvq_space.seq = tp->copied_seq;
568 tp->rcvq_space.time = tcp_time_stamp;
569}
570
571/* There is something which you must keep in mind when you analyze the
572 * behavior of the tp->ato delayed ack timeout interval. When a
573 * connection starts up, we want to ack as quickly as possible. The
574 * problem is that "good" TCP's do slow start at the beginning of data
575 * transmission. The means that until we send the first few ACK's the
576 * sender will sit on his end and only queue most of his data, because
577 * he can only send snd_cwnd unacked packets at any given time. For
578 * each ACK we send, he increments snd_cwnd and transmits more of his
579 * queue. -DaveM
580 */
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -0700581static void tcp_event_data_recv(struct sock *sk, struct sk_buff *skb)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700582{
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -0700583 struct tcp_sock *tp = tcp_sk(sk);
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700584 struct inet_connection_sock *icsk = inet_csk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700585 u32 now;
586
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700587 inet_csk_schedule_ack(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700588
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700589 tcp_measure_rcv_mss(sk, skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700590
591 tcp_rcv_rtt_measure(tp);
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +0900592
Linus Torvalds1da177e2005-04-16 15:20:36 -0700593 now = tcp_time_stamp;
594
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700595 if (!icsk->icsk_ack.ato) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700596 /* The _first_ data packet received, initialize
597 * delayed ACK engine.
598 */
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700599 tcp_incr_quickack(sk);
600 icsk->icsk_ack.ato = TCP_ATO_MIN;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700601 } else {
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700602 int m = now - icsk->icsk_ack.lrcvtime;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700603
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800604 if (m <= TCP_ATO_MIN / 2) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700605 /* The fastest case is the first. */
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700606 icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + TCP_ATO_MIN / 2;
607 } else if (m < icsk->icsk_ack.ato) {
608 icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + m;
609 if (icsk->icsk_ack.ato > icsk->icsk_rto)
610 icsk->icsk_ack.ato = icsk->icsk_rto;
611 } else if (m > icsk->icsk_rto) {
Stephen Hemmingercaa20d9a2005-11-10 17:13:47 -0800612 /* Too long gap. Apparently sender failed to
Linus Torvalds1da177e2005-04-16 15:20:36 -0700613 * restart window, so that we send ACKs quickly.
614 */
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700615 tcp_incr_quickack(sk);
Hideo Aoki3ab224b2007-12-31 00:11:19 -0800616 sk_mem_reclaim(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700617 }
618 }
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700619 icsk->icsk_ack.lrcvtime = now;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700620
621 TCP_ECN_check_ce(tp, skb);
622
623 if (skb->len >= 128)
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -0700624 tcp_grow_window(sk, skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700625}
626
Linus Torvalds1da177e2005-04-16 15:20:36 -0700627/* Called to compute a smoothed rtt estimate. The data fed to this
628 * routine either comes from timestamps, or from segments that were
629 * known _not_ to have been retransmitted [see Karn/Partridge
630 * Proceedings SIGCOMM 87]. The algorithm is from the SIGCOMM 88
631 * piece by Van Jacobson.
632 * NOTE: the next three routines used to be one big routine.
633 * To save cycles in the RFC 1323 implementation it was better to break
634 * it up into three procedures. -- erics
635 */
Stephen Hemminger2d2abba2005-11-10 16:56:12 -0800636static void tcp_rtt_estimator(struct sock *sk, const __u32 mrtt)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700637{
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -0300638 struct tcp_sock *tp = tcp_sk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700639 long m = mrtt; /* RTT */
640
Linus Torvalds1da177e2005-04-16 15:20:36 -0700641 /* The following amusing code comes from Jacobson's
642 * article in SIGCOMM '88. Note that rtt and mdev
643 * are scaled versions of rtt and mean deviation.
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +0900644 * This is designed to be as fast as possible
Linus Torvalds1da177e2005-04-16 15:20:36 -0700645 * m stands for "measurement".
646 *
647 * On a 1990 paper the rto value is changed to:
648 * RTO = rtt + 4 * mdev
649 *
650 * Funny. This algorithm seems to be very broken.
651 * These formulae increase RTO, when it should be decreased, increase
Stephen Hemminger31f34262005-11-15 15:17:10 -0800652 * too slowly, when it should be increased quickly, decrease too quickly
Linus Torvalds1da177e2005-04-16 15:20:36 -0700653 * etc. I guess in BSD RTO takes ONE value, so that it is absolutely
654 * does not matter how to _calculate_ it. Seems, it was trap
655 * that VJ failed to avoid. 8)
656 */
Stephen Hemminger2de979b2007-03-08 20:45:19 -0800657 if (m == 0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700658 m = 1;
659 if (tp->srtt != 0) {
660 m -= (tp->srtt >> 3); /* m is now error in rtt est */
661 tp->srtt += m; /* rtt = 7/8 rtt + 1/8 new */
662 if (m < 0) {
663 m = -m; /* m is now abs(error) */
664 m -= (tp->mdev >> 2); /* similar update on mdev */
665 /* This is similar to one of Eifel findings.
666 * Eifel blocks mdev updates when rtt decreases.
667 * This solution is a bit different: we use finer gain
668 * for mdev in this case (alpha*beta).
669 * Like Eifel it also prevents growth of rto,
670 * but also it limits too fast rto decreases,
671 * happening in pure Eifel.
672 */
673 if (m > 0)
674 m >>= 3;
675 } else {
676 m -= (tp->mdev >> 2); /* similar update on mdev */
677 }
678 tp->mdev += m; /* mdev = 3/4 mdev + 1/4 new */
679 if (tp->mdev > tp->mdev_max) {
680 tp->mdev_max = tp->mdev;
681 if (tp->mdev_max > tp->rttvar)
682 tp->rttvar = tp->mdev_max;
683 }
684 if (after(tp->snd_una, tp->rtt_seq)) {
685 if (tp->mdev_max < tp->rttvar)
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800686 tp->rttvar -= (tp->rttvar - tp->mdev_max) >> 2;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700687 tp->rtt_seq = tp->snd_nxt;
David S. Miller05bb1fa2007-08-30 22:10:28 -0700688 tp->mdev_max = tcp_rto_min(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700689 }
690 } else {
691 /* no previous measure. */
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800692 tp->srtt = m << 3; /* take the measured time to be rtt */
693 tp->mdev = m << 1; /* make sure rto = 3*rtt */
David S. Miller05bb1fa2007-08-30 22:10:28 -0700694 tp->mdev_max = tp->rttvar = max(tp->mdev, tcp_rto_min(sk));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700695 tp->rtt_seq = tp->snd_nxt;
696 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700697}
698
699/* Calculate rto without backoff. This is the second half of Van Jacobson's
700 * routine referred to above.
701 */
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700702static inline void tcp_set_rto(struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700703{
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700704 const struct tcp_sock *tp = tcp_sk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700705 /* Old crap is replaced with new one. 8)
706 *
707 * More seriously:
708 * 1. If rtt variance happened to be less 50msec, it is hallucination.
709 * It cannot be less due to utterly erratic ACK generation made
710 * at least by solaris and freebsd. "Erratic ACKs" has _nothing_
711 * to do with delayed acks, because at cwnd>2 true delack timeout
712 * is invisible. Actually, Linux-2.4 also generates erratic
Stephen Hemmingercaa20d9a2005-11-10 17:13:47 -0800713 * ACKs in some circumstances.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700714 */
Damian Lukowskif1ecd5d2009-08-26 00:16:31 +0000715 inet_csk(sk)->icsk_rto = __tcp_set_rto(tp);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700716
717 /* 2. Fixups made earlier cannot be right.
718 * If we do not estimate RTO correctly without them,
719 * all the algo is pure shit and should be replaced
Stephen Hemmingercaa20d9a2005-11-10 17:13:47 -0800720 * with correct one. It is exactly, which we pretend to do.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700721 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700722
Ilpo Järvinenee6aac52008-12-05 22:43:08 -0800723 /* NOTE: clamping at TCP_RTO_MIN is not required, current algo
724 * guarantees that rto is higher.
725 */
Damian Lukowskif1ecd5d2009-08-26 00:16:31 +0000726 tcp_bound_rto(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700727}
728
729/* Save metrics learned by this TCP session.
730 This function is called only, when TCP finishes successfully
731 i.e. when it enters TIME-WAIT or goes from LAST-ACK to CLOSE.
732 */
733void tcp_update_metrics(struct sock *sk)
734{
735 struct tcp_sock *tp = tcp_sk(sk);
736 struct dst_entry *dst = __sk_dst_get(sk);
737
738 if (sysctl_tcp_nometrics_save)
739 return;
740
741 dst_confirm(dst);
742
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800743 if (dst && (dst->flags & DST_HOST)) {
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -0300744 const struct inet_connection_sock *icsk = inet_csk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700745 int m;
Stephen Hemmingerc1e20f72008-07-18 23:02:15 -0700746 unsigned long rtt;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700747
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -0300748 if (icsk->icsk_backoff || !tp->srtt) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700749 /* This session failed to estimate rtt. Why?
750 * Probably, no packets returned in time.
751 * Reset our results.
752 */
753 if (!(dst_metric_locked(dst, RTAX_RTT)))
David S. Millerdefb3512010-12-08 21:16:57 -0800754 dst_metric_set(dst, RTAX_RTT, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700755 return;
756 }
757
Stephen Hemmingerc1e20f72008-07-18 23:02:15 -0700758 rtt = dst_metric_rtt(dst, RTAX_RTT);
759 m = rtt - tp->srtt;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700760
761 /* If newly calculated rtt larger than stored one,
762 * store new one. Otherwise, use EWMA. Remember,
763 * rtt overestimation is always better than underestimation.
764 */
765 if (!(dst_metric_locked(dst, RTAX_RTT))) {
766 if (m <= 0)
Stephen Hemmingerc1e20f72008-07-18 23:02:15 -0700767 set_dst_metric_rtt(dst, RTAX_RTT, tp->srtt);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700768 else
Stephen Hemmingerc1e20f72008-07-18 23:02:15 -0700769 set_dst_metric_rtt(dst, RTAX_RTT, rtt - (m >> 3));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700770 }
771
772 if (!(dst_metric_locked(dst, RTAX_RTTVAR))) {
Stephen Hemmingerc1e20f72008-07-18 23:02:15 -0700773 unsigned long var;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700774 if (m < 0)
775 m = -m;
776
777 /* Scale deviation to rttvar fixed point */
778 m >>= 1;
779 if (m < tp->mdev)
780 m = tp->mdev;
781
Stephen Hemmingerc1e20f72008-07-18 23:02:15 -0700782 var = dst_metric_rtt(dst, RTAX_RTTVAR);
783 if (m >= var)
784 var = m;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700785 else
Stephen Hemmingerc1e20f72008-07-18 23:02:15 -0700786 var -= (var - m) >> 2;
787
788 set_dst_metric_rtt(dst, RTAX_RTTVAR, var);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700789 }
790
Ilpo Järvinen0b6a05c2009-09-15 01:30:10 -0700791 if (tcp_in_initial_slowstart(tp)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700792 /* Slow start still did not finish. */
793 if (dst_metric(dst, RTAX_SSTHRESH) &&
794 !dst_metric_locked(dst, RTAX_SSTHRESH) &&
795 (tp->snd_cwnd >> 1) > dst_metric(dst, RTAX_SSTHRESH))
David S. Millerdefb3512010-12-08 21:16:57 -0800796 dst_metric_set(dst, RTAX_SSTHRESH, tp->snd_cwnd >> 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700797 if (!dst_metric_locked(dst, RTAX_CWND) &&
798 tp->snd_cwnd > dst_metric(dst, RTAX_CWND))
David S. Millerdefb3512010-12-08 21:16:57 -0800799 dst_metric_set(dst, RTAX_CWND, tp->snd_cwnd);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700800 } else if (tp->snd_cwnd > tp->snd_ssthresh &&
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -0300801 icsk->icsk_ca_state == TCP_CA_Open) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700802 /* Cong. avoidance phase, cwnd is reliable. */
803 if (!dst_metric_locked(dst, RTAX_SSTHRESH))
David S. Millerdefb3512010-12-08 21:16:57 -0800804 dst_metric_set(dst, RTAX_SSTHRESH,
805 max(tp->snd_cwnd >> 1, tp->snd_ssthresh));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700806 if (!dst_metric_locked(dst, RTAX_CWND))
David S. Millerdefb3512010-12-08 21:16:57 -0800807 dst_metric_set(dst, RTAX_CWND,
808 (dst_metric(dst, RTAX_CWND) +
809 tp->snd_cwnd) >> 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700810 } else {
811 /* Else slow start did not finish, cwnd is non-sense,
812 ssthresh may be also invalid.
813 */
814 if (!dst_metric_locked(dst, RTAX_CWND))
David S. Millerdefb3512010-12-08 21:16:57 -0800815 dst_metric_set(dst, RTAX_CWND,
816 (dst_metric(dst, RTAX_CWND) +
817 tp->snd_ssthresh) >> 1);
Satoru SATOH5ffc02a2008-05-04 22:14:42 -0700818 if (dst_metric(dst, RTAX_SSTHRESH) &&
Linus Torvalds1da177e2005-04-16 15:20:36 -0700819 !dst_metric_locked(dst, RTAX_SSTHRESH) &&
Satoru SATOH5ffc02a2008-05-04 22:14:42 -0700820 tp->snd_ssthresh > dst_metric(dst, RTAX_SSTHRESH))
David S. Millerdefb3512010-12-08 21:16:57 -0800821 dst_metric_set(dst, RTAX_SSTHRESH, tp->snd_ssthresh);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700822 }
823
824 if (!dst_metric_locked(dst, RTAX_REORDERING)) {
Satoru SATOH5ffc02a2008-05-04 22:14:42 -0700825 if (dst_metric(dst, RTAX_REORDERING) < tp->reordering &&
Linus Torvalds1da177e2005-04-16 15:20:36 -0700826 tp->reordering != sysctl_tcp_reordering)
David S. Millerdefb3512010-12-08 21:16:57 -0800827 dst_metric_set(dst, RTAX_REORDERING, tp->reordering);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700828 }
829 }
830}
831
Eric Dumazetcf533ea2011-10-21 05:22:42 -0400832__u32 tcp_init_cwnd(const struct tcp_sock *tp, const struct dst_entry *dst)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700833{
834 __u32 cwnd = (dst ? dst_metric(dst, RTAX_INITCWND) : 0);
835
Gerrit Renker22b71c82010-08-29 19:23:12 +0000836 if (!cwnd)
David S. Miller442b9632011-02-02 17:05:11 -0800837 cwnd = TCP_INIT_CWND;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700838 return min_t(__u32, cwnd, tp->snd_cwnd_clamp);
839}
840
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800841/* Set slow start threshold and cwnd not falling to slow start */
Ilpo Järvinen3cfe3ba2007-02-27 10:09:49 -0800842void tcp_enter_cwr(struct sock *sk, const int set_ssthresh)
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800843{
844 struct tcp_sock *tp = tcp_sk(sk);
Ilpo Järvinen3cfe3ba2007-02-27 10:09:49 -0800845 const struct inet_connection_sock *icsk = inet_csk(sk);
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800846
847 tp->prior_ssthresh = 0;
848 tp->bytes_acked = 0;
Ilpo Järvinene01f9d72007-03-02 13:27:25 -0800849 if (icsk->icsk_ca_state < TCP_CA_CWR) {
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800850 tp->undo_marker = 0;
Ilpo Järvinen3cfe3ba2007-02-27 10:09:49 -0800851 if (set_ssthresh)
852 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800853 tp->snd_cwnd = min(tp->snd_cwnd,
854 tcp_packets_in_flight(tp) + 1U);
855 tp->snd_cwnd_cnt = 0;
856 tp->high_seq = tp->snd_nxt;
857 tp->snd_cwnd_stamp = tcp_time_stamp;
858 TCP_ECN_queue_cwr(tp);
859
860 tcp_set_ca_state(sk, TCP_CA_CWR);
861 }
862}
863
Ilpo Järvinene60402d2007-08-09 15:14:46 +0300864/*
865 * Packet counting of FACK is based on in-order assumptions, therefore TCP
866 * disables it when reordering is detected
867 */
868static void tcp_disable_fack(struct tcp_sock *tp)
869{
Ilpo Järvinen85cc3912007-11-15 19:39:31 -0800870 /* RFC3517 uses different metric in lost marker => reset on change */
871 if (tcp_is_fack(tp))
872 tp->lost_skb_hint = NULL;
Vijay Subramanianab562222011-12-20 13:23:24 +0000873 tp->rx_opt.sack_ok &= ~TCP_FACK_ENABLED;
Ilpo Järvinene60402d2007-08-09 15:14:46 +0300874}
875
Ryousei Takano564262c2007-10-25 23:03:52 -0700876/* Take a notice that peer is sending D-SACKs */
Ilpo Järvinene60402d2007-08-09 15:14:46 +0300877static void tcp_dsack_seen(struct tcp_sock *tp)
878{
Vijay Subramanianab562222011-12-20 13:23:24 +0000879 tp->rx_opt.sack_ok |= TCP_DSACK_SEEN;
Ilpo Järvinene60402d2007-08-09 15:14:46 +0300880}
881
Linus Torvalds1da177e2005-04-16 15:20:36 -0700882/* Initialize metrics on socket. */
883
884static void tcp_init_metrics(struct sock *sk)
885{
886 struct tcp_sock *tp = tcp_sk(sk);
887 struct dst_entry *dst = __sk_dst_get(sk);
888
889 if (dst == NULL)
890 goto reset;
891
892 dst_confirm(dst);
893
894 if (dst_metric_locked(dst, RTAX_CWND))
895 tp->snd_cwnd_clamp = dst_metric(dst, RTAX_CWND);
896 if (dst_metric(dst, RTAX_SSTHRESH)) {
897 tp->snd_ssthresh = dst_metric(dst, RTAX_SSTHRESH);
898 if (tp->snd_ssthresh > tp->snd_cwnd_clamp)
899 tp->snd_ssthresh = tp->snd_cwnd_clamp;
Jerry Chu9ad7c042011-06-08 11:08:38 +0000900 } else {
901 /* ssthresh may have been reduced unnecessarily during.
902 * 3WHS. Restore it back to its initial default.
903 */
904 tp->snd_ssthresh = TCP_INFINITE_SSTHRESH;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700905 }
906 if (dst_metric(dst, RTAX_REORDERING) &&
907 tp->reordering != dst_metric(dst, RTAX_REORDERING)) {
Ilpo Järvinene60402d2007-08-09 15:14:46 +0300908 tcp_disable_fack(tp);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700909 tp->reordering = dst_metric(dst, RTAX_REORDERING);
910 }
911
Jerry Chu9ad7c042011-06-08 11:08:38 +0000912 if (dst_metric(dst, RTAX_RTT) == 0 || tp->srtt == 0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700913 goto reset;
914
915 /* Initial rtt is determined from SYN,SYN-ACK.
916 * The segment is small and rtt may appear much
917 * less than real one. Use per-dst memory
918 * to make it more realistic.
919 *
920 * A bit of theory. RTT is time passed after "normal" sized packet
Stephen Hemmingercaa20d9a2005-11-10 17:13:47 -0800921 * is sent until it is ACKed. In normal circumstances sending small
Linus Torvalds1da177e2005-04-16 15:20:36 -0700922 * packets force peer to delay ACKs and calculation is correct too.
923 * The algorithm is adaptive and, provided we follow specs, it
924 * NEVER underestimate RTT. BUT! If peer tries to make some clever
925 * tricks sort of "quick acks" for time long enough to decrease RTT
926 * to low value, and then abruptly stops to do it and starts to delay
927 * ACKs, wait for troubles.
928 */
Stephen Hemmingerc1e20f72008-07-18 23:02:15 -0700929 if (dst_metric_rtt(dst, RTAX_RTT) > tp->srtt) {
930 tp->srtt = dst_metric_rtt(dst, RTAX_RTT);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700931 tp->rtt_seq = tp->snd_nxt;
932 }
Stephen Hemmingerc1e20f72008-07-18 23:02:15 -0700933 if (dst_metric_rtt(dst, RTAX_RTTVAR) > tp->mdev) {
934 tp->mdev = dst_metric_rtt(dst, RTAX_RTTVAR);
Satoru SATOH488faa22007-12-16 14:00:19 -0800935 tp->mdev_max = tp->rttvar = max(tp->mdev, tcp_rto_min(sk));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700936 }
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700937 tcp_set_rto(sk);
Jiri Kosinad9f4fba2010-12-22 23:23:38 +0000938reset:
Jerry Chu9ad7c042011-06-08 11:08:38 +0000939 if (tp->srtt == 0) {
940 /* RFC2988bis: We've failed to get a valid RTT sample from
941 * 3WHS. This is most likely due to retransmission,
942 * including spurious one. Reset the RTO back to 3secs
943 * from the more aggressive 1sec to avoid more spurious
944 * retransmission.
Jiri Kosinad9f4fba2010-12-22 23:23:38 +0000945 */
Jerry Chu9ad7c042011-06-08 11:08:38 +0000946 tp->mdev = tp->mdev_max = tp->rttvar = TCP_TIMEOUT_FALLBACK;
947 inet_csk(sk)->icsk_rto = TCP_TIMEOUT_FALLBACK;
Jiri Kosinad9f4fba2010-12-22 23:23:38 +0000948 }
Jerry Chu9ad7c042011-06-08 11:08:38 +0000949 /* Cut cwnd down to 1 per RFC5681 if SYN or SYN-ACK has been
950 * retransmitted. In light of RFC2988bis' more aggressive 1sec
951 * initRTO, we only reset cwnd when more than 1 SYN/SYN-ACK
952 * retransmission has occurred.
953 */
954 if (tp->total_retrans > 1)
955 tp->snd_cwnd = 1;
956 else
957 tp->snd_cwnd = tcp_init_cwnd(tp, dst);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700958 tp->snd_cwnd_stamp = tcp_time_stamp;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700959}
960
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -0300961static void tcp_update_reordering(struct sock *sk, const int metric,
962 const int ts)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700963{
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -0300964 struct tcp_sock *tp = tcp_sk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700965 if (metric > tp->reordering) {
Pavel Emelyanov40b215e2008-07-03 01:05:41 -0700966 int mib_idx;
967
Linus Torvalds1da177e2005-04-16 15:20:36 -0700968 tp->reordering = min(TCP_MAX_REORDERING, metric);
969
970 /* This exciting event is worth to be remembered. 8) */
971 if (ts)
Pavel Emelyanov40b215e2008-07-03 01:05:41 -0700972 mib_idx = LINUX_MIB_TCPTSREORDER;
Ilpo Järvinene60402d2007-08-09 15:14:46 +0300973 else if (tcp_is_reno(tp))
Pavel Emelyanov40b215e2008-07-03 01:05:41 -0700974 mib_idx = LINUX_MIB_TCPRENOREORDER;
Ilpo Järvinene60402d2007-08-09 15:14:46 +0300975 else if (tcp_is_fack(tp))
Pavel Emelyanov40b215e2008-07-03 01:05:41 -0700976 mib_idx = LINUX_MIB_TCPFACKREORDER;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700977 else
Pavel Emelyanov40b215e2008-07-03 01:05:41 -0700978 mib_idx = LINUX_MIB_TCPSACKREORDER;
979
Pavel Emelyanovde0744a2008-07-16 20:31:16 -0700980 NET_INC_STATS_BH(sock_net(sk), mib_idx);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700981#if FASTRETRANS_DEBUG > 1
982 printk(KERN_DEBUG "Disorder%d %d %u f%u s%u rr%d\n",
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -0300983 tp->rx_opt.sack_ok, inet_csk(sk)->icsk_ca_state,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700984 tp->reordering,
985 tp->fackets_out,
986 tp->sacked_out,
987 tp->undo_marker ? tp->undo_retrans : 0);
988#endif
Ilpo Järvinene60402d2007-08-09 15:14:46 +0300989 tcp_disable_fack(tp);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700990 }
991}
992
Ilpo Järvinen006f5822008-09-20 21:20:20 -0700993/* This must be called before lost_out is incremented */
Ilpo Järvinenc8c213f2008-09-20 21:18:55 -0700994static void tcp_verify_retransmit_hint(struct tcp_sock *tp, struct sk_buff *skb)
995{
Ilpo Järvinen006f5822008-09-20 21:20:20 -0700996 if ((tp->retransmit_skb_hint == NULL) ||
Ilpo Järvinenc8c213f2008-09-20 21:18:55 -0700997 before(TCP_SKB_CB(skb)->seq,
998 TCP_SKB_CB(tp->retransmit_skb_hint)->seq))
Ilpo Järvinen006f5822008-09-20 21:20:20 -0700999 tp->retransmit_skb_hint = skb;
1000
1001 if (!tp->lost_out ||
1002 after(TCP_SKB_CB(skb)->end_seq, tp->retransmit_high))
1003 tp->retransmit_high = TCP_SKB_CB(skb)->end_seq;
Ilpo Järvinenc8c213f2008-09-20 21:18:55 -07001004}
1005
Ilpo Järvinen41ea36e2008-09-20 21:19:22 -07001006static void tcp_skb_mark_lost(struct tcp_sock *tp, struct sk_buff *skb)
1007{
1008 if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_ACKED))) {
1009 tcp_verify_retransmit_hint(tp, skb);
1010
1011 tp->lost_out += tcp_skb_pcount(skb);
1012 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1013 }
1014}
1015
Ilpo Järvinene1aa6802008-11-24 21:11:55 -08001016static void tcp_skb_mark_lost_uncond_verify(struct tcp_sock *tp,
1017 struct sk_buff *skb)
Ilpo Järvinen006f5822008-09-20 21:20:20 -07001018{
1019 tcp_verify_retransmit_hint(tp, skb);
1020
1021 if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_ACKED))) {
1022 tp->lost_out += tcp_skb_pcount(skb);
1023 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1024 }
1025}
1026
Linus Torvalds1da177e2005-04-16 15:20:36 -07001027/* This procedure tags the retransmission queue when SACKs arrive.
1028 *
1029 * We have three tag bits: SACKED(S), RETRANS(R) and LOST(L).
1030 * Packets in queue with these bits set are counted in variables
1031 * sacked_out, retrans_out and lost_out, correspondingly.
1032 *
1033 * Valid combinations are:
1034 * Tag InFlight Description
1035 * 0 1 - orig segment is in flight.
1036 * S 0 - nothing flies, orig reached receiver.
1037 * L 0 - nothing flies, orig lost by net.
1038 * R 2 - both orig and retransmit are in flight.
1039 * L|R 1 - orig is lost, retransmit is in flight.
1040 * S|R 1 - orig reached receiver, retrans is still in flight.
1041 * (L|S|R is logically valid, it could occur when L|R is sacked,
1042 * but it is equivalent to plain S and code short-curcuits it to S.
1043 * L|S is logically invalid, it would mean -1 packet in flight 8))
1044 *
1045 * These 6 states form finite state machine, controlled by the following events:
1046 * 1. New ACK (+SACK) arrives. (tcp_sacktag_write_queue())
1047 * 2. Retransmission. (tcp_retransmit_skb(), tcp_xmit_retransmit_queue())
Yuchung Cheng974c1232012-01-19 14:42:21 +00001048 * 3. Loss detection event of two flavors:
Linus Torvalds1da177e2005-04-16 15:20:36 -07001049 * A. Scoreboard estimator decided the packet is lost.
1050 * A'. Reno "three dupacks" marks head of queue lost.
Yuchung Cheng974c1232012-01-19 14:42:21 +00001051 * A''. Its FACK modification, head until snd.fack is lost.
1052 * B. SACK arrives sacking SND.NXT at the moment, when the
Linus Torvalds1da177e2005-04-16 15:20:36 -07001053 * segment was retransmitted.
1054 * 4. D-SACK added new rule: D-SACK changes any tag to S.
1055 *
1056 * It is pleasant to note, that state diagram turns out to be commutative,
1057 * so that we are allowed not to be bothered by order of our actions,
1058 * when multiple events arrive simultaneously. (see the function below).
1059 *
1060 * Reordering detection.
1061 * --------------------
1062 * Reordering metric is maximal distance, which a packet can be displaced
1063 * in packet stream. With SACKs we can estimate it:
1064 *
1065 * 1. SACK fills old hole and the corresponding segment was not
1066 * ever retransmitted -> reordering. Alas, we cannot use it
1067 * when segment was retransmitted.
1068 * 2. The last flaw is solved with D-SACK. D-SACK arrives
1069 * for retransmitted and already SACKed segment -> reordering..
1070 * Both of these heuristics are not used in Loss state, when we cannot
1071 * account for retransmits accurately.
Ilpo Järvinen5b3c9882007-08-24 22:54:44 -07001072 *
1073 * SACK block validation.
1074 * ----------------------
1075 *
1076 * SACK block range validation checks that the received SACK block fits to
1077 * the expected sequence limits, i.e., it is between SND.UNA and SND.NXT.
1078 * Note that SND.UNA is not included to the range though being valid because
Ilpo Järvinen0e835332007-10-01 15:28:17 -07001079 * it means that the receiver is rather inconsistent with itself reporting
1080 * SACK reneging when it should advance SND.UNA. Such SACK block this is
1081 * perfectly valid, however, in light of RFC2018 which explicitly states
1082 * that "SACK block MUST reflect the newest segment. Even if the newest
1083 * segment is going to be discarded ...", not that it looks very clever
1084 * in case of head skb. Due to potentional receiver driven attacks, we
1085 * choose to avoid immediate execution of a walk in write queue due to
1086 * reneging and defer head skb's loss recovery to standard loss recovery
1087 * procedure that will eventually trigger (nothing forbids us doing this).
Ilpo Järvinen5b3c9882007-08-24 22:54:44 -07001088 *
1089 * Implements also blockage to start_seq wrap-around. Problem lies in the
1090 * fact that though start_seq (s) is before end_seq (i.e., not reversed),
1091 * there's no guarantee that it will be before snd_nxt (n). The problem
1092 * happens when start_seq resides between end_seq wrap (e_w) and snd_nxt
1093 * wrap (s_w):
1094 *
1095 * <- outs wnd -> <- wrapzone ->
1096 * u e n u_w e_w s n_w
1097 * | | | | | | |
1098 * |<------------+------+----- TCP seqno space --------------+---------->|
1099 * ...-- <2^31 ->| |<--------...
1100 * ...---- >2^31 ------>| |<--------...
1101 *
1102 * Current code wouldn't be vulnerable but it's better still to discard such
1103 * crazy SACK blocks. Doing this check for start_seq alone closes somewhat
1104 * similar case (end_seq after snd_nxt wrap) as earlier reversed check in
1105 * snd_nxt wrap -> snd_una region will then become "well defined", i.e.,
1106 * equal to the ideal case (infinite seqno space without wrap caused issues).
1107 *
1108 * With D-SACK the lower bound is extended to cover sequence space below
1109 * SND.UNA down to undo_marker, which is the last point of interest. Yet
Ryousei Takano564262c2007-10-25 23:03:52 -07001110 * again, D-SACK block must not to go across snd_una (for the same reason as
Ilpo Järvinen5b3c9882007-08-24 22:54:44 -07001111 * for the normal SACK blocks, explained above). But there all simplicity
1112 * ends, TCP might receive valid D-SACKs below that. As long as they reside
1113 * fully below undo_marker they do not affect behavior in anyway and can
1114 * therefore be safely ignored. In rare cases (which are more or less
1115 * theoretical ones), the D-SACK will nicely cross that boundary due to skb
1116 * fragmentation and packet reordering past skb's retransmission. To consider
1117 * them correctly, the acceptable range must be extended even more though
1118 * the exact amount is rather hard to quantify. However, tp->max_window can
1119 * be used as an exaggerated estimate.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001120 */
Ilpo Järvinen5b3c9882007-08-24 22:54:44 -07001121static int tcp_is_sackblock_valid(struct tcp_sock *tp, int is_dsack,
1122 u32 start_seq, u32 end_seq)
1123{
1124 /* Too far in future, or reversed (interpretation is ambiguous) */
1125 if (after(end_seq, tp->snd_nxt) || !before(start_seq, end_seq))
1126 return 0;
1127
1128 /* Nasty start_seq wrap-around check (see comments above) */
1129 if (!before(start_seq, tp->snd_nxt))
1130 return 0;
1131
Ryousei Takano564262c2007-10-25 23:03:52 -07001132 /* In outstanding window? ...This is valid exit for D-SACKs too.
Ilpo Järvinen5b3c9882007-08-24 22:54:44 -07001133 * start_seq == snd_una is non-sensical (see comments above)
1134 */
1135 if (after(start_seq, tp->snd_una))
1136 return 1;
1137
1138 if (!is_dsack || !tp->undo_marker)
1139 return 0;
1140
1141 /* ...Then it's D-SACK, and must reside below snd_una completely */
Zheng Yanf779b2d2011-09-18 22:37:34 -04001142 if (after(end_seq, tp->snd_una))
Ilpo Järvinen5b3c9882007-08-24 22:54:44 -07001143 return 0;
1144
1145 if (!before(start_seq, tp->undo_marker))
1146 return 1;
1147
1148 /* Too old */
1149 if (!after(end_seq, tp->undo_marker))
1150 return 0;
1151
1152 /* Undo_marker boundary crossing (overestimates a lot). Known already:
1153 * start_seq < undo_marker and end_seq >= undo_marker.
1154 */
1155 return !before(start_seq, end_seq - tp->max_window);
1156}
1157
Ilpo Järvinen1c1e87e2007-10-10 02:45:32 -07001158/* Check for lost retransmit. This superb idea is borrowed from "ratehalving".
Yuchung Cheng974c1232012-01-19 14:42:21 +00001159 * Event "B". Later note: FACK people cheated me again 8), we have to account
Ilpo Järvinen1c1e87e2007-10-10 02:45:32 -07001160 * for reordering! Ugly, but should help.
Ilpo Järvinenf785a8e2007-10-11 17:35:41 -07001161 *
1162 * Search retransmitted skbs from write_queue that were sent when snd_nxt was
1163 * less than what is now known to be received by the other end (derived from
Ilpo Järvinen9f58f3b2007-11-15 19:42:54 -08001164 * highest SACK block). Also calculate the lowest snd_nxt among the remaining
1165 * retransmitted skbs to avoid some costly processing per ACKs.
Ilpo Järvinen1c1e87e2007-10-10 02:45:32 -07001166 */
Ilpo Järvinen407ef1d2007-12-02 00:47:57 +02001167static void tcp_mark_lost_retrans(struct sock *sk)
Ilpo Järvinen1c1e87e2007-10-10 02:45:32 -07001168{
Ilpo Järvinen9f58f3b2007-11-15 19:42:54 -08001169 const struct inet_connection_sock *icsk = inet_csk(sk);
Ilpo Järvinen1c1e87e2007-10-10 02:45:32 -07001170 struct tcp_sock *tp = tcp_sk(sk);
1171 struct sk_buff *skb;
Ilpo Järvinenf785a8e2007-10-11 17:35:41 -07001172 int cnt = 0;
Ilpo Järvinendf2e0142007-10-18 05:07:57 -07001173 u32 new_low_seq = tp->snd_nxt;
Ilpo Järvinen6859d492007-12-02 00:48:06 +02001174 u32 received_upto = tcp_highest_sack_seq(tp);
Ilpo Järvinen9f58f3b2007-11-15 19:42:54 -08001175
1176 if (!tcp_is_fack(tp) || !tp->retrans_out ||
1177 !after(received_upto, tp->lost_retrans_low) ||
1178 icsk->icsk_ca_state != TCP_CA_Recovery)
Ilpo Järvinen407ef1d2007-12-02 00:47:57 +02001179 return;
Ilpo Järvinen1c1e87e2007-10-10 02:45:32 -07001180
1181 tcp_for_write_queue(skb, sk) {
1182 u32 ack_seq = TCP_SKB_CB(skb)->ack_seq;
1183
1184 if (skb == tcp_send_head(sk))
1185 break;
Ilpo Järvinenf785a8e2007-10-11 17:35:41 -07001186 if (cnt == tp->retrans_out)
Ilpo Järvinen1c1e87e2007-10-10 02:45:32 -07001187 break;
1188 if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
1189 continue;
1190
Ilpo Järvinenf785a8e2007-10-11 17:35:41 -07001191 if (!(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS))
1192 continue;
1193
Ilpo Järvinend0af4162009-02-28 04:44:32 +00001194 /* TODO: We would like to get rid of tcp_is_fack(tp) only
1195 * constraint here (see above) but figuring out that at
1196 * least tp->reordering SACK blocks reside between ack_seq
1197 * and received_upto is not easy task to do cheaply with
1198 * the available datastructures.
1199 *
1200 * Whether FACK should check here for tp->reordering segs
1201 * in-between one could argue for either way (it would be
1202 * rather simple to implement as we could count fack_count
1203 * during the walk and do tp->fackets_out - fack_count).
1204 */
1205 if (after(received_upto, ack_seq)) {
Ilpo Järvinen1c1e87e2007-10-10 02:45:32 -07001206 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
1207 tp->retrans_out -= tcp_skb_pcount(skb);
1208
Ilpo Järvinen006f5822008-09-20 21:20:20 -07001209 tcp_skb_mark_lost_uncond_verify(tp, skb);
Pavel Emelyanovde0744a2008-07-16 20:31:16 -07001210 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPLOSTRETRANSMIT);
Ilpo Järvinenf785a8e2007-10-11 17:35:41 -07001211 } else {
Ilpo Järvinendf2e0142007-10-18 05:07:57 -07001212 if (before(ack_seq, new_low_seq))
Ilpo Järvinenb08d6cb2007-10-11 17:36:13 -07001213 new_low_seq = ack_seq;
Ilpo Järvinenf785a8e2007-10-11 17:35:41 -07001214 cnt += tcp_skb_pcount(skb);
Ilpo Järvinen1c1e87e2007-10-10 02:45:32 -07001215 }
1216 }
Ilpo Järvinenb08d6cb2007-10-11 17:36:13 -07001217
1218 if (tp->retrans_out)
1219 tp->lost_retrans_low = new_low_seq;
Ilpo Järvinen1c1e87e2007-10-10 02:45:32 -07001220}
Ilpo Järvinen5b3c9882007-08-24 22:54:44 -07001221
Eric Dumazetcf533ea2011-10-21 05:22:42 -04001222static int tcp_check_dsack(struct sock *sk, const struct sk_buff *ack_skb,
David S. Millerd06e0212007-06-18 22:43:06 -07001223 struct tcp_sack_block_wire *sp, int num_sacks,
1224 u32 prior_snd_una)
1225{
Pavel Emelyanov1ed83462008-07-16 20:29:51 -07001226 struct tcp_sock *tp = tcp_sk(sk);
Harvey Harrisond3e2ce32008-05-02 16:26:16 -07001227 u32 start_seq_0 = get_unaligned_be32(&sp[0].start_seq);
1228 u32 end_seq_0 = get_unaligned_be32(&sp[0].end_seq);
David S. Millerd06e0212007-06-18 22:43:06 -07001229 int dup_sack = 0;
1230
1231 if (before(start_seq_0, TCP_SKB_CB(ack_skb)->ack_seq)) {
1232 dup_sack = 1;
Ilpo Järvinene60402d2007-08-09 15:14:46 +03001233 tcp_dsack_seen(tp);
Pavel Emelyanovde0744a2008-07-16 20:31:16 -07001234 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPDSACKRECV);
David S. Millerd06e0212007-06-18 22:43:06 -07001235 } else if (num_sacks > 1) {
Harvey Harrisond3e2ce32008-05-02 16:26:16 -07001236 u32 end_seq_1 = get_unaligned_be32(&sp[1].end_seq);
1237 u32 start_seq_1 = get_unaligned_be32(&sp[1].start_seq);
David S. Millerd06e0212007-06-18 22:43:06 -07001238
1239 if (!after(end_seq_0, end_seq_1) &&
1240 !before(start_seq_0, start_seq_1)) {
1241 dup_sack = 1;
Ilpo Järvinene60402d2007-08-09 15:14:46 +03001242 tcp_dsack_seen(tp);
Pavel Emelyanovde0744a2008-07-16 20:31:16 -07001243 NET_INC_STATS_BH(sock_net(sk),
1244 LINUX_MIB_TCPDSACKOFORECV);
David S. Millerd06e0212007-06-18 22:43:06 -07001245 }
1246 }
1247
1248 /* D-SACK for already forgotten data... Do dumb counting. */
Yuchung Chengc24f6912011-02-07 12:57:04 +00001249 if (dup_sack && tp->undo_marker && tp->undo_retrans &&
David S. Millerd06e0212007-06-18 22:43:06 -07001250 !after(end_seq_0, prior_snd_una) &&
1251 after(end_seq_0, tp->undo_marker))
1252 tp->undo_retrans--;
1253
1254 return dup_sack;
1255}
1256
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001257struct tcp_sacktag_state {
1258 int reord;
1259 int fack_count;
1260 int flag;
1261};
1262
Ilpo Järvinend1935942007-10-11 17:34:25 -07001263/* Check if skb is fully within the SACK block. In presence of GSO skbs,
1264 * the incoming SACK may not exactly match but we can find smaller MSS
1265 * aligned portion of it that matches. Therefore we might need to fragment
1266 * which may fail and creates some hassle (caller must handle error case
1267 * returns).
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001268 *
1269 * FIXME: this could be merged to shift decision code
Ilpo Järvinend1935942007-10-11 17:34:25 -07001270 */
Adrian Bunk0f79efd2007-10-26 03:57:36 -07001271static int tcp_match_skb_to_sack(struct sock *sk, struct sk_buff *skb,
1272 u32 start_seq, u32 end_seq)
Ilpo Järvinend1935942007-10-11 17:34:25 -07001273{
1274 int in_sack, err;
1275 unsigned int pkt_len;
Ilpo Järvinenadb92db2008-11-24 21:13:50 -08001276 unsigned int mss;
Ilpo Järvinend1935942007-10-11 17:34:25 -07001277
1278 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
1279 !before(end_seq, TCP_SKB_CB(skb)->end_seq);
1280
1281 if (tcp_skb_pcount(skb) > 1 && !in_sack &&
1282 after(TCP_SKB_CB(skb)->end_seq, start_seq)) {
Ilpo Järvinenadb92db2008-11-24 21:13:50 -08001283 mss = tcp_skb_mss(skb);
Ilpo Järvinend1935942007-10-11 17:34:25 -07001284 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq);
1285
Ilpo Järvinenadb92db2008-11-24 21:13:50 -08001286 if (!in_sack) {
Ilpo Järvinend1935942007-10-11 17:34:25 -07001287 pkt_len = start_seq - TCP_SKB_CB(skb)->seq;
Ilpo Järvinenadb92db2008-11-24 21:13:50 -08001288 if (pkt_len < mss)
1289 pkt_len = mss;
1290 } else {
Ilpo Järvinend1935942007-10-11 17:34:25 -07001291 pkt_len = end_seq - TCP_SKB_CB(skb)->seq;
Ilpo Järvinenadb92db2008-11-24 21:13:50 -08001292 if (pkt_len < mss)
1293 return -EINVAL;
1294 }
1295
1296 /* Round if necessary so that SACKs cover only full MSSes
1297 * and/or the remaining small portion (if present)
1298 */
1299 if (pkt_len > mss) {
1300 unsigned int new_len = (pkt_len / mss) * mss;
1301 if (!in_sack && new_len < pkt_len) {
1302 new_len += mss;
1303 if (new_len > skb->len)
1304 return 0;
1305 }
1306 pkt_len = new_len;
1307 }
1308 err = tcp_fragment(sk, skb, pkt_len, mss);
Ilpo Järvinend1935942007-10-11 17:34:25 -07001309 if (err < 0)
1310 return err;
1311 }
1312
1313 return in_sack;
1314}
1315
Neal Cardwellcc9a6722012-02-12 18:37:09 +00001316/* Mark the given newly-SACKed range as such, adjusting counters and hints. */
1317static u8 tcp_sacktag_one(struct sock *sk,
1318 struct tcp_sacktag_state *state, u8 sacked,
1319 u32 start_seq, u32 end_seq,
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001320 int dup_sack, int pcount)
Ilpo Järvinen9e10c472007-11-15 19:44:56 -08001321{
Ilpo Järvinen6859d492007-12-02 00:48:06 +02001322 struct tcp_sock *tp = tcp_sk(sk);
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001323 int fack_count = state->fack_count;
Ilpo Järvinen9e10c472007-11-15 19:44:56 -08001324
1325 /* Account D-SACK for retransmitted packet. */
1326 if (dup_sack && (sacked & TCPCB_RETRANS)) {
Yuchung Chengc24f6912011-02-07 12:57:04 +00001327 if (tp->undo_marker && tp->undo_retrans &&
Neal Cardwellcc9a6722012-02-12 18:37:09 +00001328 after(end_seq, tp->undo_marker))
Ilpo Järvinen9e10c472007-11-15 19:44:56 -08001329 tp->undo_retrans--;
Ilpo Järvinenede9f3b2007-12-02 00:47:58 +02001330 if (sacked & TCPCB_SACKED_ACKED)
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001331 state->reord = min(fack_count, state->reord);
Ilpo Järvinen9e10c472007-11-15 19:44:56 -08001332 }
1333
1334 /* Nothing to do; acked frame is about to be dropped (was ACKed). */
Neal Cardwellcc9a6722012-02-12 18:37:09 +00001335 if (!after(end_seq, tp->snd_una))
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001336 return sacked;
Ilpo Järvinen9e10c472007-11-15 19:44:56 -08001337
1338 if (!(sacked & TCPCB_SACKED_ACKED)) {
1339 if (sacked & TCPCB_SACKED_RETRANS) {
1340 /* If the segment is not tagged as lost,
1341 * we do not clear RETRANS, believing
1342 * that retransmission is still in flight.
1343 */
1344 if (sacked & TCPCB_LOST) {
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001345 sacked &= ~(TCPCB_LOST|TCPCB_SACKED_RETRANS);
Ilpo Järvinenf58b22f2008-11-24 21:14:43 -08001346 tp->lost_out -= pcount;
1347 tp->retrans_out -= pcount;
Ilpo Järvinen9e10c472007-11-15 19:44:56 -08001348 }
1349 } else {
1350 if (!(sacked & TCPCB_RETRANS)) {
1351 /* New sack for not retransmitted frame,
1352 * which was in hole. It is reordering.
1353 */
Neal Cardwellcc9a6722012-02-12 18:37:09 +00001354 if (before(start_seq,
Ilpo Järvinen9e10c472007-11-15 19:44:56 -08001355 tcp_highest_sack_seq(tp)))
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001356 state->reord = min(fack_count,
1357 state->reord);
Ilpo Järvinen9e10c472007-11-15 19:44:56 -08001358
1359 /* SACK enhanced F-RTO (RFC4138; Appendix B) */
Neal Cardwellcc9a6722012-02-12 18:37:09 +00001360 if (!after(end_seq, tp->frto_highmark))
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001361 state->flag |= FLAG_ONLY_ORIG_SACKED;
Ilpo Järvinen9e10c472007-11-15 19:44:56 -08001362 }
1363
1364 if (sacked & TCPCB_LOST) {
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001365 sacked &= ~TCPCB_LOST;
Ilpo Järvinenf58b22f2008-11-24 21:14:43 -08001366 tp->lost_out -= pcount;
Ilpo Järvinen9e10c472007-11-15 19:44:56 -08001367 }
1368 }
1369
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001370 sacked |= TCPCB_SACKED_ACKED;
1371 state->flag |= FLAG_DATA_SACKED;
Ilpo Järvinenf58b22f2008-11-24 21:14:43 -08001372 tp->sacked_out += pcount;
Ilpo Järvinen9e10c472007-11-15 19:44:56 -08001373
Ilpo Järvinenf58b22f2008-11-24 21:14:43 -08001374 fack_count += pcount;
Ilpo Järvinen9e10c472007-11-15 19:44:56 -08001375
1376 /* Lost marker hint past SACKed? Tweak RFC3517 cnt */
1377 if (!tcp_is_fack(tp) && (tp->lost_skb_hint != NULL) &&
Neal Cardwellcc9a6722012-02-12 18:37:09 +00001378 before(start_seq, TCP_SKB_CB(tp->lost_skb_hint)->seq))
Ilpo Järvinenf58b22f2008-11-24 21:14:43 -08001379 tp->lost_cnt_hint += pcount;
Ilpo Järvinen9e10c472007-11-15 19:44:56 -08001380
1381 if (fack_count > tp->fackets_out)
1382 tp->fackets_out = fack_count;
Ilpo Järvinen9e10c472007-11-15 19:44:56 -08001383 }
1384
1385 /* D-SACK. We can detect redundant retransmission in S|R and plain R
1386 * frames and clear it. undo_retrans is decreased above, L|R frames
1387 * are accounted above as well.
1388 */
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001389 if (dup_sack && (sacked & TCPCB_SACKED_RETRANS)) {
1390 sacked &= ~TCPCB_SACKED_RETRANS;
Ilpo Järvinenf58b22f2008-11-24 21:14:43 -08001391 tp->retrans_out -= pcount;
Ilpo Järvinen9e10c472007-11-15 19:44:56 -08001392 }
1393
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001394 return sacked;
Ilpo Järvinen9e10c472007-11-15 19:44:56 -08001395}
1396
Neal Cardwelldaef52b2012-02-12 18:37:10 +00001397/* Shift newly-SACKed bytes from this skb to the immediately previous
1398 * already-SACKed sk_buff. Mark the newly-SACKed bytes as such.
1399 */
Ilpo Järvinen50133162008-12-05 22:42:41 -08001400static int tcp_shifted_skb(struct sock *sk, struct sk_buff *skb,
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001401 struct tcp_sacktag_state *state,
Ilpo Järvinen9ec06ff2009-03-01 00:21:36 -08001402 unsigned int pcount, int shifted, int mss,
1403 int dup_sack)
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001404{
1405 struct tcp_sock *tp = tcp_sk(sk);
Ilpo Järvinen50133162008-12-05 22:42:41 -08001406 struct sk_buff *prev = tcp_write_queue_prev(sk, skb);
Neal Cardwelldaef52b2012-02-12 18:37:10 +00001407 u32 start_seq = TCP_SKB_CB(skb)->seq; /* start of newly-SACKed */
1408 u32 end_seq = start_seq + shifted; /* end of newly-SACKed */
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001409
1410 BUG_ON(!pcount);
1411
Neal Cardwell4c90d3b2012-02-26 10:06:19 +00001412 /* Adjust counters and hints for the newly sacked sequence
1413 * range but discard the return value since prev is already
1414 * marked. We must tag the range first because the seq
1415 * advancement below implicitly advances
1416 * tcp_highest_sack_seq() when skb is highest_sack.
1417 */
1418 tcp_sacktag_one(sk, state, TCP_SKB_CB(skb)->sacked,
1419 start_seq, end_seq, dup_sack, pcount);
1420
1421 if (skb == tp->lost_skb_hint)
Neal Cardwell0af2a0d2012-02-13 20:22:08 +00001422 tp->lost_cnt_hint += pcount;
1423
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001424 TCP_SKB_CB(prev)->end_seq += shifted;
1425 TCP_SKB_CB(skb)->seq += shifted;
1426
1427 skb_shinfo(prev)->gso_segs += pcount;
1428 BUG_ON(skb_shinfo(skb)->gso_segs < pcount);
1429 skb_shinfo(skb)->gso_segs -= pcount;
1430
1431 /* When we're adding to gso_segs == 1, gso_size will be zero,
1432 * in theory this shouldn't be necessary but as long as DSACK
1433 * code can come after this skb later on it's better to keep
1434 * setting gso_size to something.
1435 */
1436 if (!skb_shinfo(prev)->gso_size) {
1437 skb_shinfo(prev)->gso_size = mss;
1438 skb_shinfo(prev)->gso_type = sk->sk_gso_type;
1439 }
1440
1441 /* CHECKME: To clear or not to clear? Mimics normal skb currently */
1442 if (skb_shinfo(skb)->gso_segs <= 1) {
1443 skb_shinfo(skb)->gso_size = 0;
1444 skb_shinfo(skb)->gso_type = 0;
1445 }
1446
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001447 /* Difference in this won't matter, both ACKed by the same cumul. ACK */
1448 TCP_SKB_CB(prev)->sacked |= (TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS);
1449
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001450 if (skb->len > 0) {
1451 BUG_ON(!tcp_skb_pcount(skb));
Ilpo Järvinen111cc8b2008-11-24 21:27:22 -08001452 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_SACKSHIFTED);
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001453 return 0;
1454 }
1455
1456 /* Whole SKB was eaten :-) */
1457
Ilpo Järvinen92ee76b2008-11-24 21:26:56 -08001458 if (skb == tp->retransmit_skb_hint)
1459 tp->retransmit_skb_hint = prev;
1460 if (skb == tp->scoreboard_skb_hint)
1461 tp->scoreboard_skb_hint = prev;
1462 if (skb == tp->lost_skb_hint) {
1463 tp->lost_skb_hint = prev;
1464 tp->lost_cnt_hint -= tcp_skb_pcount(prev);
1465 }
1466
Eric Dumazet4de075e2011-09-27 13:25:05 -04001467 TCP_SKB_CB(skb)->tcp_flags |= TCP_SKB_CB(prev)->tcp_flags;
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001468 if (skb == tcp_highest_sack(sk))
1469 tcp_advance_highest_sack(sk, skb);
1470
1471 tcp_unlink_write_queue(skb, sk);
1472 sk_wmem_free_skb(sk, skb);
1473
Ilpo Järvinen111cc8b2008-11-24 21:27:22 -08001474 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_SACKMERGED);
1475
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001476 return 1;
1477}
1478
1479/* I wish gso_size would have a bit more sane initialization than
1480 * something-or-zero which complicates things
1481 */
Eric Dumazetcf533ea2011-10-21 05:22:42 -04001482static int tcp_skb_seglen(const struct sk_buff *skb)
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001483{
Ilpo Järvinen775ffab2008-12-05 22:41:26 -08001484 return tcp_skb_pcount(skb) == 1 ? skb->len : tcp_skb_mss(skb);
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001485}
1486
1487/* Shifting pages past head area doesn't work */
Eric Dumazetcf533ea2011-10-21 05:22:42 -04001488static int skb_can_shift(const struct sk_buff *skb)
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001489{
1490 return !skb_headlen(skb) && skb_is_nonlinear(skb);
1491}
1492
1493/* Try collapsing SACK blocks spanning across multiple skbs to a single
1494 * skb.
1495 */
1496static struct sk_buff *tcp_shift_skb_data(struct sock *sk, struct sk_buff *skb,
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001497 struct tcp_sacktag_state *state,
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001498 u32 start_seq, u32 end_seq,
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001499 int dup_sack)
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001500{
1501 struct tcp_sock *tp = tcp_sk(sk);
1502 struct sk_buff *prev;
1503 int mss;
1504 int pcount = 0;
1505 int len;
1506 int in_sack;
1507
1508 if (!sk_can_gso(sk))
1509 goto fallback;
1510
1511 /* Normally R but no L won't result in plain S */
1512 if (!dup_sack &&
Ilpo Järvinen9969ca52008-12-05 22:41:06 -08001513 (TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_RETRANS)) == TCPCB_SACKED_RETRANS)
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001514 goto fallback;
1515 if (!skb_can_shift(skb))
1516 goto fallback;
1517 /* This frame is about to be dropped (was ACKed). */
1518 if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
1519 goto fallback;
1520
1521 /* Can only happen with delayed DSACK + discard craziness */
1522 if (unlikely(skb == tcp_write_queue_head(sk)))
1523 goto fallback;
1524 prev = tcp_write_queue_prev(sk, skb);
1525
1526 if ((TCP_SKB_CB(prev)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED)
1527 goto fallback;
1528
1529 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
1530 !before(end_seq, TCP_SKB_CB(skb)->end_seq);
1531
1532 if (in_sack) {
1533 len = skb->len;
1534 pcount = tcp_skb_pcount(skb);
Ilpo Järvinen775ffab2008-12-05 22:41:26 -08001535 mss = tcp_skb_seglen(skb);
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001536
1537 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1538 * drop this restriction as unnecessary
1539 */
Ilpo Järvinen775ffab2008-12-05 22:41:26 -08001540 if (mss != tcp_skb_seglen(prev))
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001541 goto fallback;
1542 } else {
1543 if (!after(TCP_SKB_CB(skb)->end_seq, start_seq))
1544 goto noop;
1545 /* CHECKME: This is non-MSS split case only?, this will
1546 * cause skipped skbs due to advancing loop btw, original
1547 * has that feature too
1548 */
1549 if (tcp_skb_pcount(skb) <= 1)
1550 goto noop;
1551
1552 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq);
1553 if (!in_sack) {
1554 /* TODO: head merge to next could be attempted here
1555 * if (!after(TCP_SKB_CB(skb)->end_seq, end_seq)),
1556 * though it might not be worth of the additional hassle
1557 *
1558 * ...we can probably just fallback to what was done
1559 * previously. We could try merging non-SACKed ones
1560 * as well but it probably isn't going to buy off
1561 * because later SACKs might again split them, and
1562 * it would make skb timestamp tracking considerably
1563 * harder problem.
1564 */
1565 goto fallback;
1566 }
1567
1568 len = end_seq - TCP_SKB_CB(skb)->seq;
1569 BUG_ON(len < 0);
1570 BUG_ON(len > skb->len);
1571
1572 /* MSS boundaries should be honoured or else pcount will
1573 * severely break even though it makes things bit trickier.
1574 * Optimize common case to avoid most of the divides
1575 */
1576 mss = tcp_skb_mss(skb);
1577
1578 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1579 * drop this restriction as unnecessary
1580 */
Ilpo Järvinen775ffab2008-12-05 22:41:26 -08001581 if (mss != tcp_skb_seglen(prev))
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001582 goto fallback;
1583
1584 if (len == mss) {
1585 pcount = 1;
1586 } else if (len < mss) {
1587 goto noop;
1588 } else {
1589 pcount = len / mss;
1590 len = pcount * mss;
1591 }
1592 }
1593
Neal Cardwell4648dc92012-03-05 19:35:04 +00001594 /* tcp_sacktag_one() won't SACK-tag ranges below snd_una */
1595 if (!after(TCP_SKB_CB(skb)->seq + len, tp->snd_una))
1596 goto fallback;
1597
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001598 if (!skb_shift(prev, skb, len))
1599 goto fallback;
Ilpo Järvinen9ec06ff2009-03-01 00:21:36 -08001600 if (!tcp_shifted_skb(sk, skb, state, pcount, len, mss, dup_sack))
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001601 goto out;
1602
1603 /* Hole filled allows collapsing with the next as well, this is very
1604 * useful when hole on every nth skb pattern happens
1605 */
1606 if (prev == tcp_write_queue_tail(sk))
1607 goto out;
1608 skb = tcp_write_queue_next(sk, prev);
1609
Ilpo Järvinenf0bc52f2008-12-05 22:40:47 -08001610 if (!skb_can_shift(skb) ||
1611 (skb == tcp_send_head(sk)) ||
1612 ((TCP_SKB_CB(skb)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED) ||
Ilpo Järvinen775ffab2008-12-05 22:41:26 -08001613 (mss != tcp_skb_seglen(skb)))
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001614 goto out;
1615
1616 len = skb->len;
1617 if (skb_shift(prev, skb, len)) {
1618 pcount += tcp_skb_pcount(skb);
Ilpo Järvinen9ec06ff2009-03-01 00:21:36 -08001619 tcp_shifted_skb(sk, skb, state, tcp_skb_pcount(skb), len, mss, 0);
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001620 }
1621
1622out:
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001623 state->fack_count += pcount;
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001624 return prev;
1625
1626noop:
1627 return skb;
1628
1629fallback:
Ilpo Järvinen111cc8b2008-11-24 21:27:22 -08001630 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_SACKSHIFTFALLBACK);
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001631 return NULL;
1632}
1633
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001634static struct sk_buff *tcp_sacktag_walk(struct sk_buff *skb, struct sock *sk,
1635 struct tcp_sack_block *next_dup,
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001636 struct tcp_sacktag_state *state,
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001637 u32 start_seq, u32 end_seq,
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001638 int dup_sack_in)
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001639{
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001640 struct tcp_sock *tp = tcp_sk(sk);
1641 struct sk_buff *tmp;
1642
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001643 tcp_for_write_queue_from(skb, sk) {
1644 int in_sack = 0;
1645 int dup_sack = dup_sack_in;
1646
1647 if (skb == tcp_send_head(sk))
1648 break;
1649
1650 /* queue is in-order => we can short-circuit the walk early */
1651 if (!before(TCP_SKB_CB(skb)->seq, end_seq))
1652 break;
1653
1654 if ((next_dup != NULL) &&
1655 before(TCP_SKB_CB(skb)->seq, next_dup->end_seq)) {
1656 in_sack = tcp_match_skb_to_sack(sk, skb,
1657 next_dup->start_seq,
1658 next_dup->end_seq);
1659 if (in_sack > 0)
1660 dup_sack = 1;
1661 }
1662
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001663 /* skb reference here is a bit tricky to get right, since
1664 * shifting can eat and free both this skb and the next,
1665 * so not even _safe variant of the loop is enough.
1666 */
1667 if (in_sack <= 0) {
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001668 tmp = tcp_shift_skb_data(sk, skb, state,
1669 start_seq, end_seq, dup_sack);
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001670 if (tmp != NULL) {
1671 if (tmp != skb) {
1672 skb = tmp;
1673 continue;
1674 }
1675
1676 in_sack = 0;
1677 } else {
1678 in_sack = tcp_match_skb_to_sack(sk, skb,
1679 start_seq,
1680 end_seq);
1681 }
1682 }
1683
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001684 if (unlikely(in_sack < 0))
1685 break;
1686
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001687 if (in_sack) {
Neal Cardwellcc9a6722012-02-12 18:37:09 +00001688 TCP_SKB_CB(skb)->sacked =
1689 tcp_sacktag_one(sk,
1690 state,
1691 TCP_SKB_CB(skb)->sacked,
1692 TCP_SKB_CB(skb)->seq,
1693 TCP_SKB_CB(skb)->end_seq,
1694 dup_sack,
1695 tcp_skb_pcount(skb));
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001696
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001697 if (!before(TCP_SKB_CB(skb)->seq,
1698 tcp_highest_sack_seq(tp)))
1699 tcp_advance_highest_sack(sk, skb);
1700 }
1701
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001702 state->fack_count += tcp_skb_pcount(skb);
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001703 }
1704 return skb;
1705}
1706
1707/* Avoid all extra work that is being done by sacktag while walking in
1708 * a normal way
1709 */
1710static struct sk_buff *tcp_sacktag_skip(struct sk_buff *skb, struct sock *sk,
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001711 struct tcp_sacktag_state *state,
1712 u32 skip_to_seq)
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001713{
1714 tcp_for_write_queue_from(skb, sk) {
1715 if (skb == tcp_send_head(sk))
1716 break;
1717
Ilpo Järvinene8bae272008-11-24 21:12:28 -08001718 if (after(TCP_SKB_CB(skb)->end_seq, skip_to_seq))
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001719 break;
Ilpo Järvinend152a7d2008-03-03 12:10:16 -08001720
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001721 state->fack_count += tcp_skb_pcount(skb);
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001722 }
1723 return skb;
1724}
1725
1726static struct sk_buff *tcp_maybe_skipping_dsack(struct sk_buff *skb,
1727 struct sock *sk,
1728 struct tcp_sack_block *next_dup,
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001729 struct tcp_sacktag_state *state,
1730 u32 skip_to_seq)
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001731{
1732 if (next_dup == NULL)
1733 return skb;
1734
1735 if (before(next_dup->start_seq, skip_to_seq)) {
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001736 skb = tcp_sacktag_skip(skb, sk, state, next_dup->start_seq);
1737 skb = tcp_sacktag_walk(skb, sk, NULL, state,
1738 next_dup->start_seq, next_dup->end_seq,
1739 1);
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001740 }
1741
1742 return skb;
1743}
1744
Eric Dumazetcf533ea2011-10-21 05:22:42 -04001745static int tcp_sack_cache_ok(const struct tcp_sock *tp, const struct tcp_sack_block *cache)
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001746{
1747 return cache < tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache);
1748}
1749
Linus Torvalds1da177e2005-04-16 15:20:36 -07001750static int
Eric Dumazetcf533ea2011-10-21 05:22:42 -04001751tcp_sacktag_write_queue(struct sock *sk, const struct sk_buff *ack_skb,
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001752 u32 prior_snd_una)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001753{
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03001754 const struct inet_connection_sock *icsk = inet_csk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001755 struct tcp_sock *tp = tcp_sk(sk);
Eric Dumazetcf533ea2011-10-21 05:22:42 -04001756 const unsigned char *ptr = (skb_transport_header(ack_skb) +
1757 TCP_SKB_CB(ack_skb)->sacked);
Ilpo Järvinenfd6dad62007-11-15 19:49:47 -08001758 struct tcp_sack_block_wire *sp_wire = (struct tcp_sack_block_wire *)(ptr+2);
Adam Langley4389dde2008-07-19 00:07:02 -07001759 struct tcp_sack_block sp[TCP_NUM_SACKS];
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001760 struct tcp_sack_block *cache;
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001761 struct tcp_sacktag_state state;
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001762 struct sk_buff *skb;
Adam Langley4389dde2008-07-19 00:07:02 -07001763 int num_sacks = min(TCP_NUM_SACKS, (ptr[1] - TCPOLEN_SACK_BASE) >> 3);
Ilpo Järvinenfd6dad62007-11-15 19:49:47 -08001764 int used_sacks;
Ilpo Järvinen7769f402007-06-15 15:14:04 -07001765 int found_dup_sack = 0;
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001766 int i, j;
Baruch Evenfda03fb2007-02-04 23:35:57 -08001767 int first_sack_index;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001768
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001769 state.flag = 0;
1770 state.reord = tp->packets_out;
1771
Ilpo Järvinend738cd82007-03-24 21:03:23 -07001772 if (!tp->sacked_out) {
Ilpo Järvinende83c052007-10-07 23:37:55 -07001773 if (WARN_ON(tp->fackets_out))
1774 tp->fackets_out = 0;
Ilpo Järvinen6859d492007-12-02 00:48:06 +02001775 tcp_highest_sack_reset(sk);
Ilpo Järvinend738cd82007-03-24 21:03:23 -07001776 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001777
Pavel Emelyanov1ed83462008-07-16 20:29:51 -07001778 found_dup_sack = tcp_check_dsack(sk, ack_skb, sp_wire,
David S. Millerd06e0212007-06-18 22:43:06 -07001779 num_sacks, prior_snd_una);
1780 if (found_dup_sack)
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001781 state.flag |= FLAG_DSACKING_ACK;
Baruch Even6f746512007-02-04 23:36:42 -08001782
1783 /* Eliminate too old ACKs, but take into
1784 * account more or less fresh ones, they can
1785 * contain valid SACK info.
1786 */
1787 if (before(TCP_SKB_CB(ack_skb)->ack_seq, prior_snd_una - tp->max_window))
1788 return 0;
1789
Ilpo Järvinen96a2d412007-11-14 15:47:18 -08001790 if (!tp->packets_out)
1791 goto out;
1792
Ilpo Järvinenfd6dad62007-11-15 19:49:47 -08001793 used_sacks = 0;
1794 first_sack_index = 0;
1795 for (i = 0; i < num_sacks; i++) {
1796 int dup_sack = !i && found_dup_sack;
1797
Harvey Harrisond3e2ce32008-05-02 16:26:16 -07001798 sp[used_sacks].start_seq = get_unaligned_be32(&sp_wire[i].start_seq);
1799 sp[used_sacks].end_seq = get_unaligned_be32(&sp_wire[i].end_seq);
Ilpo Järvinenfd6dad62007-11-15 19:49:47 -08001800
1801 if (!tcp_is_sackblock_valid(tp, dup_sack,
1802 sp[used_sacks].start_seq,
1803 sp[used_sacks].end_seq)) {
Pavel Emelyanov40b215e2008-07-03 01:05:41 -07001804 int mib_idx;
1805
Ilpo Järvinenfd6dad62007-11-15 19:49:47 -08001806 if (dup_sack) {
1807 if (!tp->undo_marker)
Pavel Emelyanov40b215e2008-07-03 01:05:41 -07001808 mib_idx = LINUX_MIB_TCPDSACKIGNOREDNOUNDO;
Ilpo Järvinenfd6dad62007-11-15 19:49:47 -08001809 else
Pavel Emelyanov40b215e2008-07-03 01:05:41 -07001810 mib_idx = LINUX_MIB_TCPDSACKIGNOREDOLD;
Ilpo Järvinenfd6dad62007-11-15 19:49:47 -08001811 } else {
1812 /* Don't count olds caused by ACK reordering */
1813 if ((TCP_SKB_CB(ack_skb)->ack_seq != tp->snd_una) &&
1814 !after(sp[used_sacks].end_seq, tp->snd_una))
1815 continue;
Pavel Emelyanov40b215e2008-07-03 01:05:41 -07001816 mib_idx = LINUX_MIB_TCPSACKDISCARD;
Ilpo Järvinenfd6dad62007-11-15 19:49:47 -08001817 }
Pavel Emelyanov40b215e2008-07-03 01:05:41 -07001818
Pavel Emelyanovde0744a2008-07-16 20:31:16 -07001819 NET_INC_STATS_BH(sock_net(sk), mib_idx);
Ilpo Järvinenfd6dad62007-11-15 19:49:47 -08001820 if (i == 0)
1821 first_sack_index = -1;
1822 continue;
1823 }
1824
1825 /* Ignore very old stuff early */
1826 if (!after(sp[used_sacks].end_seq, prior_snd_una))
1827 continue;
1828
1829 used_sacks++;
1830 }
1831
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001832 /* order SACK blocks to allow in order walk of the retrans queue */
1833 for (i = used_sacks - 1; i > 0; i--) {
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001834 for (j = 0; j < i; j++) {
1835 if (after(sp[j].start_seq, sp[j + 1].start_seq)) {
Ilpo Järvinena0bffff2009-03-21 13:36:17 -07001836 swap(sp[j], sp[j + 1]);
Stephen Hemminger6a438bb2005-11-10 17:14:59 -08001837
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001838 /* Track where the first SACK block goes to */
1839 if (j == first_sack_index)
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001840 first_sack_index = j + 1;
Stephen Hemminger6a438bb2005-11-10 17:14:59 -08001841 }
1842 }
1843 }
1844
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001845 skb = tcp_write_queue_head(sk);
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001846 state.fack_count = 0;
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001847 i = 0;
1848
1849 if (!tp->sacked_out) {
1850 /* It's already past, so skip checking against it */
1851 cache = tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache);
1852 } else {
1853 cache = tp->recv_sack_cache;
1854 /* Skip empty blocks in at head of the cache */
1855 while (tcp_sack_cache_ok(tp, cache) && !cache->start_seq &&
1856 !cache->end_seq)
1857 cache++;
Baruch Evenfda03fb2007-02-04 23:35:57 -08001858 }
1859
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001860 while (i < used_sacks) {
Ilpo Järvinenfd6dad62007-11-15 19:49:47 -08001861 u32 start_seq = sp[i].start_seq;
1862 u32 end_seq = sp[i].end_seq;
Ilpo Järvinen7769f402007-06-15 15:14:04 -07001863 int dup_sack = (found_dup_sack && (i == first_sack_index));
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001864 struct tcp_sack_block *next_dup = NULL;
Ilpo Järvinene56d6cd2007-11-01 00:09:37 -07001865
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001866 if (found_dup_sack && ((i + 1) == first_sack_index))
1867 next_dup = &sp[i + 1];
Linus Torvalds1da177e2005-04-16 15:20:36 -07001868
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001869 /* Skip too early cached blocks */
1870 while (tcp_sack_cache_ok(tp, cache) &&
1871 !before(start_seq, cache->end_seq))
1872 cache++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001873
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001874 /* Can skip some work by looking recv_sack_cache? */
1875 if (tcp_sack_cache_ok(tp, cache) && !dup_sack &&
1876 after(end_seq, cache->start_seq)) {
David S. Millerfe067e82007-03-07 12:12:44 -08001877
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001878 /* Head todo? */
1879 if (before(start_seq, cache->start_seq)) {
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001880 skb = tcp_sacktag_skip(skb, sk, &state,
1881 start_seq);
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001882 skb = tcp_sacktag_walk(skb, sk, next_dup,
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001883 &state,
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001884 start_seq,
1885 cache->start_seq,
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001886 dup_sack);
Baruch Evenfda03fb2007-02-04 23:35:57 -08001887 }
Stephen Hemminger6a438bb2005-11-10 17:14:59 -08001888
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001889 /* Rest of the block already fully processed? */
Ilpo Järvinen20de20b2007-11-16 16:17:05 -08001890 if (!after(end_seq, cache->end_seq))
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001891 goto advance_sp;
Ilpo Järvinen20de20b2007-11-16 16:17:05 -08001892
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001893 skb = tcp_maybe_skipping_dsack(skb, sk, next_dup,
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001894 &state,
1895 cache->end_seq);
Ilpo Järvinene56d6cd2007-11-01 00:09:37 -07001896
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001897 /* ...tail remains todo... */
Ilpo Järvinen6859d492007-12-02 00:48:06 +02001898 if (tcp_highest_sack_seq(tp) == cache->end_seq) {
Ilpo Järvinen20de20b2007-11-16 16:17:05 -08001899 /* ...but better entrypoint exists! */
Ilpo Järvinen6859d492007-12-02 00:48:06 +02001900 skb = tcp_highest_sack(sk);
1901 if (skb == NULL)
1902 break;
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001903 state.fack_count = tp->fackets_out;
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001904 cache++;
1905 goto walk;
1906 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001907
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001908 skb = tcp_sacktag_skip(skb, sk, &state, cache->end_seq);
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001909 /* Check overlap against next cached too (past this one already) */
1910 cache++;
1911 continue;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001912 }
Ilpo Järvinenfbd52eb2007-11-10 21:24:19 -08001913
Ilpo Järvinen6859d492007-12-02 00:48:06 +02001914 if (!before(start_seq, tcp_highest_sack_seq(tp))) {
1915 skb = tcp_highest_sack(sk);
1916 if (skb == NULL)
1917 break;
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001918 state.fack_count = tp->fackets_out;
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001919 }
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001920 skb = tcp_sacktag_skip(skb, sk, &state, start_seq);
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001921
1922walk:
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001923 skb = tcp_sacktag_walk(skb, sk, next_dup, &state,
1924 start_seq, end_seq, dup_sack);
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001925
1926advance_sp:
Ilpo Järvinenfbd52eb2007-11-10 21:24:19 -08001927 /* SACK enhanced FRTO (RFC4138, Appendix B): Clearing correct
1928 * due to in-order walk
1929 */
1930 if (after(end_seq, tp->frto_highmark))
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001931 state.flag &= ~FLAG_ONLY_ORIG_SACKED;
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001932
1933 i++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001934 }
1935
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001936 /* Clear the head of the cache sack blocks so we can skip it next time */
1937 for (i = 0; i < ARRAY_SIZE(tp->recv_sack_cache) - used_sacks; i++) {
1938 tp->recv_sack_cache[i].start_seq = 0;
1939 tp->recv_sack_cache[i].end_seq = 0;
1940 }
1941 for (j = 0; j < used_sacks; j++)
1942 tp->recv_sack_cache[i++] = sp[j];
1943
Ilpo Järvinen407ef1d2007-12-02 00:47:57 +02001944 tcp_mark_lost_retrans(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001945
Ilpo Järvinen86426c22007-08-09 14:45:17 +03001946 tcp_verify_left_out(tp);
1947
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001948 if ((state.reord < tp->fackets_out) &&
Ilpo Järvinenf5771112007-11-15 19:35:11 -08001949 ((icsk->icsk_ca_state != TCP_CA_Loss) || tp->undo_marker) &&
Ilpo Järvinenc5e7af02007-02-23 16:22:06 -08001950 (!tp->frto_highmark || after(tp->snd_una, tp->frto_highmark)))
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001951 tcp_update_reordering(sk, tp->fackets_out - state.reord, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001952
Ilpo Järvinen96a2d412007-11-14 15:47:18 -08001953out:
1954
Linus Torvalds1da177e2005-04-16 15:20:36 -07001955#if FASTRETRANS_DEBUG > 0
Ilpo Järvinen547b7922008-07-25 21:43:18 -07001956 WARN_ON((int)tp->sacked_out < 0);
1957 WARN_ON((int)tp->lost_out < 0);
1958 WARN_ON((int)tp->retrans_out < 0);
1959 WARN_ON((int)tcp_packets_in_flight(tp) < 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001960#endif
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001961 return state.flag;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001962}
1963
Ilpo Järvinen882beba2008-04-07 22:33:07 -07001964/* Limits sacked_out so that sum with lost_out isn't ever larger than
1965 * packets_out. Returns zero if sacked_out adjustement wasn't necessary.
Ilpo Järvinen30935cf2007-02-21 23:01:36 -08001966 */
Ilpo Järvinen8eecaba2008-11-25 13:45:29 -08001967static int tcp_limit_reno_sacked(struct tcp_sock *tp)
Ilpo Järvinen4ddf6672007-05-27 01:52:00 -07001968{
Ilpo Järvinen4ddf6672007-05-27 01:52:00 -07001969 u32 holes;
1970
1971 holes = max(tp->lost_out, 1U);
1972 holes = min(holes, tp->packets_out);
1973
1974 if ((tp->sacked_out + holes) > tp->packets_out) {
1975 tp->sacked_out = tp->packets_out - holes;
Ilpo Järvinen882beba2008-04-07 22:33:07 -07001976 return 1;
Ilpo Järvinen4ddf6672007-05-27 01:52:00 -07001977 }
Ilpo Järvinen882beba2008-04-07 22:33:07 -07001978 return 0;
1979}
1980
1981/* If we receive more dupacks than we expected counting segments
1982 * in assumption of absent reordering, interpret this as reordering.
1983 * The only another reason could be bug in receiver TCP.
1984 */
1985static void tcp_check_reno_reordering(struct sock *sk, const int addend)
1986{
1987 struct tcp_sock *tp = tcp_sk(sk);
1988 if (tcp_limit_reno_sacked(tp))
1989 tcp_update_reordering(sk, tp->packets_out + addend, 0);
Ilpo Järvinen4ddf6672007-05-27 01:52:00 -07001990}
1991
1992/* Emulate SACKs for SACKless connection: account for a new dupack. */
1993
1994static void tcp_add_reno_sack(struct sock *sk)
1995{
1996 struct tcp_sock *tp = tcp_sk(sk);
1997 tp->sacked_out++;
1998 tcp_check_reno_reordering(sk, 0);
Ilpo Järvinen005903b2007-08-09 14:44:16 +03001999 tcp_verify_left_out(tp);
Ilpo Järvinen4ddf6672007-05-27 01:52:00 -07002000}
2001
2002/* Account for ACK, ACKing some data in Reno Recovery phase. */
2003
2004static void tcp_remove_reno_sacks(struct sock *sk, int acked)
2005{
2006 struct tcp_sock *tp = tcp_sk(sk);
2007
2008 if (acked > 0) {
2009 /* One ACK acked hole. The rest eat duplicate ACKs. */
Ilpo Järvinen056834d2007-12-31 14:57:14 -08002010 if (acked - 1 >= tp->sacked_out)
Ilpo Järvinen4ddf6672007-05-27 01:52:00 -07002011 tp->sacked_out = 0;
2012 else
Ilpo Järvinen056834d2007-12-31 14:57:14 -08002013 tp->sacked_out -= acked - 1;
Ilpo Järvinen4ddf6672007-05-27 01:52:00 -07002014 }
2015 tcp_check_reno_reordering(sk, acked);
Ilpo Järvinen005903b2007-08-09 14:44:16 +03002016 tcp_verify_left_out(tp);
Ilpo Järvinen4ddf6672007-05-27 01:52:00 -07002017}
2018
2019static inline void tcp_reset_reno_sack(struct tcp_sock *tp)
2020{
2021 tp->sacked_out = 0;
Ilpo Järvinen4ddf6672007-05-27 01:52:00 -07002022}
2023
Ilpo Järvinen62ab2222008-05-08 01:09:11 -07002024static int tcp_is_sackfrto(const struct tcp_sock *tp)
2025{
2026 return (sysctl_tcp_frto == 0x2) && !tcp_is_reno(tp);
2027}
2028
Ilpo Järvinen95eacd22007-10-01 15:27:42 -07002029/* F-RTO can only be used if TCP has never retransmitted anything other than
2030 * head (SACK enhanced variant from Appendix B of RFC4138 is more robust here)
2031 */
Ilpo Järvinen46d0de42007-02-21 23:10:39 -08002032int tcp_use_frto(struct sock *sk)
Ilpo Järvinenbdaae172007-02-21 22:59:58 -08002033{
2034 const struct tcp_sock *tp = tcp_sk(sk);
Ilpo Järvinen6adb4f72008-04-07 22:33:57 -07002035 const struct inet_connection_sock *icsk = inet_csk(sk);
Ilpo Järvinen46d0de42007-02-21 23:10:39 -08002036 struct sk_buff *skb;
Ilpo Järvinenbdaae172007-02-21 22:59:58 -08002037
Ilpo Järvinen575ee712007-04-30 00:39:55 -07002038 if (!sysctl_tcp_frto)
Ilpo Järvinen46d0de42007-02-21 23:10:39 -08002039 return 0;
2040
Ilpo Järvinen6adb4f72008-04-07 22:33:57 -07002041 /* MTU probe and F-RTO won't really play nicely along currently */
2042 if (icsk->icsk_mtup.probe_size)
2043 return 0;
2044
Ilpo Järvinen62ab2222008-05-08 01:09:11 -07002045 if (tcp_is_sackfrto(tp))
Ilpo Järvinen4dc26652007-02-21 23:16:11 -08002046 return 1;
2047
Ilpo Järvinen46d0de42007-02-21 23:10:39 -08002048 /* Avoid expensive walking of rexmit queue if possible */
2049 if (tp->retrans_out > 1)
2050 return 0;
2051
David S. Millerfe067e82007-03-07 12:12:44 -08002052 skb = tcp_write_queue_head(sk);
David S. Miller28e34872008-09-23 02:51:41 -07002053 if (tcp_skb_is_last(sk, skb))
2054 return 1;
David S. Millerfe067e82007-03-07 12:12:44 -08002055 skb = tcp_write_queue_next(sk, skb); /* Skips head */
2056 tcp_for_write_queue_from(skb, sk) {
2057 if (skb == tcp_send_head(sk))
2058 break;
Ilpo Järvinen056834d2007-12-31 14:57:14 -08002059 if (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS)
Ilpo Järvinen46d0de42007-02-21 23:10:39 -08002060 return 0;
2061 /* Short-circuit when first non-SACKed skb has been checked */
Ilpo Järvinen056834d2007-12-31 14:57:14 -08002062 if (!(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
Ilpo Järvinen46d0de42007-02-21 23:10:39 -08002063 break;
2064 }
2065 return 1;
Ilpo Järvinenbdaae172007-02-21 22:59:58 -08002066}
2067
Ilpo Järvinen30935cf2007-02-21 23:01:36 -08002068/* RTO occurred, but do not yet enter Loss state. Instead, defer RTO
2069 * recovery a bit and use heuristics in tcp_process_frto() to detect if
Ilpo Järvinend1a54c62007-02-21 23:11:57 -08002070 * the RTO was spurious. Only clear SACKED_RETRANS of the head here to
2071 * keep retrans_out counting accurate (with SACK F-RTO, other than head
2072 * may still have that bit set); TCPCB_LOST and remaining SACKED_RETRANS
2073 * bits are handled if the Loss state is really to be entered (in
2074 * tcp_enter_frto_loss).
Ilpo Järvinen7487c482007-02-21 23:02:30 -08002075 *
2076 * Do like tcp_enter_loss() would; when RTO expires the second time it
2077 * does:
2078 * "Reduce ssthresh if it has not yet been made inside this window."
Linus Torvalds1da177e2005-04-16 15:20:36 -07002079 */
2080void tcp_enter_frto(struct sock *sk)
2081{
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002082 const struct inet_connection_sock *icsk = inet_csk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002083 struct tcp_sock *tp = tcp_sk(sk);
2084 struct sk_buff *skb;
2085
Ilpo Järvinen7487c482007-02-21 23:02:30 -08002086 if ((!tp->frto_counter && icsk->icsk_ca_state <= TCP_CA_Disorder) ||
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09002087 tp->snd_una == tp->high_seq ||
Ilpo Järvinen7487c482007-02-21 23:02:30 -08002088 ((icsk->icsk_ca_state == TCP_CA_Loss || tp->frto_counter) &&
2089 !icsk->icsk_retransmits)) {
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002090 tp->prior_ssthresh = tcp_current_ssthresh(sk);
Ilpo Järvinen66e93e42007-02-21 23:13:47 -08002091 /* Our state is too optimistic in ssthresh() call because cwnd
Ryousei Takano564262c2007-10-25 23:03:52 -07002092 * is not reduced until tcp_enter_frto_loss() when previous F-RTO
Ilpo Järvinen66e93e42007-02-21 23:13:47 -08002093 * recovery has not yet completed. Pattern would be this: RTO,
2094 * Cumulative ACK, RTO (2xRTO for the same segment does not end
2095 * up here twice).
2096 * RFC4138 should be more specific on what to do, even though
2097 * RTO is quite unlikely to occur after the first Cumulative ACK
2098 * due to back-off and complexity of triggering events ...
2099 */
2100 if (tp->frto_counter) {
2101 u32 stored_cwnd;
2102 stored_cwnd = tp->snd_cwnd;
2103 tp->snd_cwnd = 2;
2104 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
2105 tp->snd_cwnd = stored_cwnd;
2106 } else {
2107 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
2108 }
2109 /* ... in theory, cong.control module could do "any tricks" in
2110 * ssthresh(), which means that ca_state, lost bits and lost_out
2111 * counter would have to be faked before the call occurs. We
2112 * consider that too expensive, unlikely and hacky, so modules
2113 * using these in ssthresh() must deal these incompatibility
2114 * issues if they receives CA_EVENT_FRTO and frto_counter != 0
2115 */
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002116 tcp_ca_event(sk, CA_EVENT_FRTO);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002117 }
2118
Linus Torvalds1da177e2005-04-16 15:20:36 -07002119 tp->undo_marker = tp->snd_una;
2120 tp->undo_retrans = 0;
2121
David S. Millerfe067e82007-03-07 12:12:44 -08002122 skb = tcp_write_queue_head(sk);
Ilpo Järvinen009a2e32007-09-20 11:34:38 -07002123 if (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS)
2124 tp->undo_marker = 0;
Ilpo Järvinend1a54c62007-02-21 23:11:57 -08002125 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
Ilpo Järvinen522e7542007-02-21 22:54:52 -08002126 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
Ilpo Järvinend1a54c62007-02-21 23:11:57 -08002127 tp->retrans_out -= tcp_skb_pcount(skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002128 }
Ilpo Järvinen005903b2007-08-09 14:44:16 +03002129 tcp_verify_left_out(tp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002130
Ilpo Järvinen746aa322007-11-13 21:01:23 -08002131 /* Too bad if TCP was application limited */
2132 tp->snd_cwnd = min(tp->snd_cwnd, tcp_packets_in_flight(tp) + 1);
2133
Ilpo Järvinen4dc26652007-02-21 23:16:11 -08002134 /* Earlier loss recovery underway (see RFC4138; Appendix B).
2135 * The last condition is necessary at least in tp->frto_counter case.
2136 */
Ilpo Järvinen62ab2222008-05-08 01:09:11 -07002137 if (tcp_is_sackfrto(tp) && (tp->frto_counter ||
Ilpo Järvinen4dc26652007-02-21 23:16:11 -08002138 ((1 << icsk->icsk_ca_state) & (TCPF_CA_Recovery|TCPF_CA_Loss))) &&
2139 after(tp->high_seq, tp->snd_una)) {
2140 tp->frto_highmark = tp->high_seq;
2141 } else {
2142 tp->frto_highmark = tp->snd_nxt;
2143 }
Ilpo Järvinen7b0eb222007-02-21 23:03:35 -08002144 tcp_set_ca_state(sk, TCP_CA_Disorder);
2145 tp->high_seq = tp->snd_nxt;
Ilpo Järvinen7487c482007-02-21 23:02:30 -08002146 tp->frto_counter = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002147}
2148
2149/* Enter Loss state after F-RTO was applied. Dupack arrived after RTO,
2150 * which indicates that we should follow the traditional RTO recovery,
2151 * i.e. mark everything lost and do go-back-N retransmission.
2152 */
Ilpo Järvinend1a54c62007-02-21 23:11:57 -08002153static void tcp_enter_frto_loss(struct sock *sk, int allowed_segments, int flag)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002154{
2155 struct tcp_sock *tp = tcp_sk(sk);
2156 struct sk_buff *skb;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002157
Linus Torvalds1da177e2005-04-16 15:20:36 -07002158 tp->lost_out = 0;
Ilpo Järvinend1a54c62007-02-21 23:11:57 -08002159 tp->retrans_out = 0;
Ilpo Järvinene60402d2007-08-09 15:14:46 +03002160 if (tcp_is_reno(tp))
Ilpo Järvinen9bff40f2007-05-27 01:53:49 -07002161 tcp_reset_reno_sack(tp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002162
David S. Millerfe067e82007-03-07 12:12:44 -08002163 tcp_for_write_queue(skb, sk) {
2164 if (skb == tcp_send_head(sk))
2165 break;
Ilpo Järvinen23aeeec2007-11-13 21:03:13 -08002166
2167 TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
Ilpo Järvinend1a54c62007-02-21 23:11:57 -08002168 /*
2169 * Count the retransmission made on RTO correctly (only when
2170 * waiting for the first ACK and did not get it)...
2171 */
Ilpo Järvinen056834d2007-12-31 14:57:14 -08002172 if ((tp->frto_counter == 1) && !(flag & FLAG_DATA_ACKED)) {
Ilpo Järvinen0a9f2a42007-07-15 00:19:29 -07002173 /* For some reason this R-bit might get cleared? */
2174 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS)
2175 tp->retrans_out += tcp_skb_pcount(skb);
Ilpo Järvinend1a54c62007-02-21 23:11:57 -08002176 /* ...enter this if branch just for the first segment */
2177 flag |= FLAG_DATA_ACKED;
2178 } else {
Ilpo Järvinen009a2e32007-09-20 11:34:38 -07002179 if (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS)
2180 tp->undo_marker = 0;
Ilpo Järvinen23aeeec2007-11-13 21:03:13 -08002181 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
Ilpo Järvinend1a54c62007-02-21 23:11:57 -08002182 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002183
Ilpo Järvinen79d445162008-05-13 02:54:19 -07002184 /* Marking forward transmissions that were made after RTO lost
2185 * can cause unnecessary retransmissions in some scenarios,
2186 * SACK blocks will mitigate that in some but not in all cases.
2187 * We used to not mark them but it was causing break-ups with
2188 * receivers that do only in-order receival.
2189 *
2190 * TODO: we could detect presence of such receiver and select
2191 * different behavior per flow.
2192 */
2193 if (!(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) {
Ilpo Järvinen9bff40f2007-05-27 01:53:49 -07002194 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
2195 tp->lost_out += tcp_skb_pcount(skb);
Ilpo Järvinen006f5822008-09-20 21:20:20 -07002196 tp->retransmit_high = TCP_SKB_CB(skb)->end_seq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002197 }
2198 }
Ilpo Järvinen005903b2007-08-09 14:44:16 +03002199 tcp_verify_left_out(tp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002200
Ilpo Järvinen95c49222007-02-21 23:05:18 -08002201 tp->snd_cwnd = tcp_packets_in_flight(tp) + allowed_segments;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002202 tp->snd_cwnd_cnt = 0;
2203 tp->snd_cwnd_stamp = tcp_time_stamp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002204 tp->frto_counter = 0;
Ilpo Järvinen16e90682007-10-11 17:32:31 -07002205 tp->bytes_acked = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002206
2207 tp->reordering = min_t(unsigned int, tp->reordering,
Ilpo Järvinen056834d2007-12-31 14:57:14 -08002208 sysctl_tcp_reordering);
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002209 tcp_set_ca_state(sk, TCP_CA_Loss);
Ilpo Järvinen79d445162008-05-13 02:54:19 -07002210 tp->high_seq = tp->snd_nxt;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002211 TCP_ECN_queue_cwr(tp);
Stephen Hemminger6a438bb2005-11-10 17:14:59 -08002212
Ilpo Järvinen64edc272008-09-20 21:18:32 -07002213 tcp_clear_all_retrans_hints(tp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002214}
2215
Ilpo Järvinen4cd82992007-10-11 17:34:57 -07002216static void tcp_clear_retrans_partial(struct tcp_sock *tp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002217{
Linus Torvalds1da177e2005-04-16 15:20:36 -07002218 tp->retrans_out = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002219 tp->lost_out = 0;
2220
2221 tp->undo_marker = 0;
2222 tp->undo_retrans = 0;
2223}
2224
Ilpo Järvinen4cd82992007-10-11 17:34:57 -07002225void tcp_clear_retrans(struct tcp_sock *tp)
2226{
2227 tcp_clear_retrans_partial(tp);
2228
2229 tp->fackets_out = 0;
2230 tp->sacked_out = 0;
2231}
2232
Linus Torvalds1da177e2005-04-16 15:20:36 -07002233/* Enter Loss state. If "how" is not zero, forget all SACK information
2234 * and reset tags completely, otherwise preserve SACKs. If receiver
2235 * dropped its ofo queue, we will know this due to reneging detection.
2236 */
2237void tcp_enter_loss(struct sock *sk, int how)
2238{
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002239 const struct inet_connection_sock *icsk = inet_csk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002240 struct tcp_sock *tp = tcp_sk(sk);
2241 struct sk_buff *skb;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002242
2243 /* Reduce ssthresh if it has not yet been made inside this window. */
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002244 if (icsk->icsk_ca_state <= TCP_CA_Disorder || tp->snd_una == tp->high_seq ||
2245 (icsk->icsk_ca_state == TCP_CA_Loss && !icsk->icsk_retransmits)) {
2246 tp->prior_ssthresh = tcp_current_ssthresh(sk);
2247 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
2248 tcp_ca_event(sk, CA_EVENT_LOSS);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002249 }
2250 tp->snd_cwnd = 1;
2251 tp->snd_cwnd_cnt = 0;
2252 tp->snd_cwnd_stamp = tcp_time_stamp;
2253
Stephen Hemminger9772efb2005-11-10 17:09:53 -08002254 tp->bytes_acked = 0;
Ilpo Järvinen4cd82992007-10-11 17:34:57 -07002255 tcp_clear_retrans_partial(tp);
2256
2257 if (tcp_is_reno(tp))
2258 tcp_reset_reno_sack(tp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002259
Ilpo Järvinenb7689202007-09-20 11:37:19 -07002260 if (!how) {
2261 /* Push undo marker, if it was plain RTO and nothing
2262 * was retransmitted. */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002263 tp->undo_marker = tp->snd_una;
Ilpo Järvinenb7689202007-09-20 11:37:19 -07002264 } else {
Ilpo Järvinen4cd82992007-10-11 17:34:57 -07002265 tp->sacked_out = 0;
2266 tp->fackets_out = 0;
Ilpo Järvinenb7689202007-09-20 11:37:19 -07002267 }
Ilpo Järvinen64edc272008-09-20 21:18:32 -07002268 tcp_clear_all_retrans_hints(tp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002269
David S. Millerfe067e82007-03-07 12:12:44 -08002270 tcp_for_write_queue(skb, sk) {
2271 if (skb == tcp_send_head(sk))
2272 break;
Ilpo Järvinen4cd82992007-10-11 17:34:57 -07002273
Ilpo Järvinen056834d2007-12-31 14:57:14 -08002274 if (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002275 tp->undo_marker = 0;
2276 TCP_SKB_CB(skb)->sacked &= (~TCPCB_TAGBITS)|TCPCB_SACKED_ACKED;
2277 if (!(TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_ACKED) || how) {
2278 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_ACKED;
2279 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
2280 tp->lost_out += tcp_skb_pcount(skb);
Ilpo Järvinen006f5822008-09-20 21:20:20 -07002281 tp->retransmit_high = TCP_SKB_CB(skb)->end_seq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002282 }
2283 }
Ilpo Järvinen005903b2007-08-09 14:44:16 +03002284 tcp_verify_left_out(tp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002285
2286 tp->reordering = min_t(unsigned int, tp->reordering,
Ilpo Järvinen056834d2007-12-31 14:57:14 -08002287 sysctl_tcp_reordering);
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002288 tcp_set_ca_state(sk, TCP_CA_Loss);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002289 tp->high_seq = tp->snd_nxt;
2290 TCP_ECN_queue_cwr(tp);
Ryousei Takano564262c2007-10-25 23:03:52 -07002291 /* Abort F-RTO algorithm if one is in progress */
Ilpo Järvinen580e5722007-05-19 13:56:57 -07002292 tp->frto_counter = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002293}
2294
Ilpo Järvinencadbd032007-12-31 04:49:21 -08002295/* If ACK arrived pointing to a remembered SACK, it means that our
2296 * remembered SACKs do not reflect real state of receiver i.e.
2297 * receiver _host_ is heavily congested (or buggy).
2298 *
2299 * Do processing similar to RTO timeout.
2300 */
2301static int tcp_check_sack_reneging(struct sock *sk, int flag)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002302{
Ilpo Järvinencadbd032007-12-31 04:49:21 -08002303 if (flag & FLAG_SACK_RENEGING) {
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002304 struct inet_connection_sock *icsk = inet_csk(sk);
Pavel Emelyanovde0744a2008-07-16 20:31:16 -07002305 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPSACKRENEGING);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002306
2307 tcp_enter_loss(sk, 1);
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002308 icsk->icsk_retransmits++;
David S. Millerfe067e82007-03-07 12:12:44 -08002309 tcp_retransmit_skb(sk, tcp_write_queue_head(sk));
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07002310 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002311 icsk->icsk_rto, TCP_RTO_MAX);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002312 return 1;
2313 }
2314 return 0;
2315}
2316
Eric Dumazetcf533ea2011-10-21 05:22:42 -04002317static inline int tcp_fackets_out(const struct tcp_sock *tp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002318{
Ilpo Järvinen056834d2007-12-31 14:57:14 -08002319 return tcp_is_reno(tp) ? tp->sacked_out + 1 : tp->fackets_out;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002320}
2321
Ilpo Järvinen85cc3912007-11-15 19:39:31 -08002322/* Heurestics to calculate number of duplicate ACKs. There's no dupACKs
2323 * counter when SACK is enabled (without SACK, sacked_out is used for
2324 * that purpose).
2325 *
2326 * Instead, with FACK TCP uses fackets_out that includes both SACKed
2327 * segments up to the highest received SACK block so far and holes in
2328 * between them.
2329 *
2330 * With reordering, holes may still be in flight, so RFC3517 recovery
2331 * uses pure sacked_out (total number of SACKed segments) even though
2332 * it violates the RFC that uses duplicate ACKs, often these are equal
2333 * but when e.g. out-of-window ACKs or packet duplication occurs,
2334 * they differ. Since neither occurs due to loss, TCP should really
2335 * ignore them.
2336 */
Eric Dumazetcf533ea2011-10-21 05:22:42 -04002337static inline int tcp_dupack_heuristics(const struct tcp_sock *tp)
Ilpo Järvinen85cc3912007-11-15 19:39:31 -08002338{
2339 return tcp_is_fack(tp) ? tp->fackets_out : tp->sacked_out + 1;
2340}
2341
Eric Dumazetcf533ea2011-10-21 05:22:42 -04002342static inline int tcp_skb_timedout(const struct sock *sk,
2343 const struct sk_buff *skb)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002344{
Eric Dumazeta02cec22010-09-22 20:43:57 +00002345 return tcp_time_stamp - TCP_SKB_CB(skb)->when > inet_csk(sk)->icsk_rto;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002346}
2347
Eric Dumazetcf533ea2011-10-21 05:22:42 -04002348static inline int tcp_head_timedout(const struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002349{
Eric Dumazetcf533ea2011-10-21 05:22:42 -04002350 const struct tcp_sock *tp = tcp_sk(sk);
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002351
Linus Torvalds1da177e2005-04-16 15:20:36 -07002352 return tp->packets_out &&
David S. Millerfe067e82007-03-07 12:12:44 -08002353 tcp_skb_timedout(sk, tcp_write_queue_head(sk));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002354}
2355
2356/* Linux NewReno/SACK/FACK/ECN state machine.
2357 * --------------------------------------
2358 *
2359 * "Open" Normal state, no dubious events, fast path.
2360 * "Disorder" In all the respects it is "Open",
2361 * but requires a bit more attention. It is entered when
2362 * we see some SACKs or dupacks. It is split of "Open"
2363 * mainly to move some processing from fast path to slow one.
2364 * "CWR" CWND was reduced due to some Congestion Notification event.
2365 * It can be ECN, ICMP source quench, local device congestion.
2366 * "Recovery" CWND was reduced, we are fast-retransmitting.
2367 * "Loss" CWND was reduced due to RTO timeout or SACK reneging.
2368 *
2369 * tcp_fastretrans_alert() is entered:
2370 * - each incoming ACK, if state is not "Open"
2371 * - when arrived ACK is unusual, namely:
2372 * * SACK
2373 * * Duplicate ACK.
2374 * * ECN ECE.
2375 *
2376 * Counting packets in flight is pretty simple.
2377 *
2378 * in_flight = packets_out - left_out + retrans_out
2379 *
2380 * packets_out is SND.NXT-SND.UNA counted in packets.
2381 *
2382 * retrans_out is number of retransmitted segments.
2383 *
2384 * left_out is number of segments left network, but not ACKed yet.
2385 *
2386 * left_out = sacked_out + lost_out
2387 *
2388 * sacked_out: Packets, which arrived to receiver out of order
2389 * and hence not ACKed. With SACKs this number is simply
2390 * amount of SACKed data. Even without SACKs
2391 * it is easy to give pretty reliable estimate of this number,
2392 * counting duplicate ACKs.
2393 *
2394 * lost_out: Packets lost by network. TCP has no explicit
2395 * "loss notification" feedback from network (for now).
2396 * It means that this number can be only _guessed_.
2397 * Actually, it is the heuristics to predict lossage that
2398 * distinguishes different algorithms.
2399 *
2400 * F.e. after RTO, when all the queue is considered as lost,
2401 * lost_out = packets_out and in_flight = retrans_out.
2402 *
2403 * Essentially, we have now two algorithms counting
2404 * lost packets.
2405 *
2406 * FACK: It is the simplest heuristics. As soon as we decided
2407 * that something is lost, we decide that _all_ not SACKed
2408 * packets until the most forward SACK are lost. I.e.
2409 * lost_out = fackets_out - sacked_out and left_out = fackets_out.
2410 * It is absolutely correct estimate, if network does not reorder
2411 * packets. And it loses any connection to reality when reordering
2412 * takes place. We use FACK by default until reordering
2413 * is suspected on the path to this destination.
2414 *
2415 * NewReno: when Recovery is entered, we assume that one segment
2416 * is lost (classic Reno). While we are in Recovery and
2417 * a partial ACK arrives, we assume that one more packet
2418 * is lost (NewReno). This heuristics are the same in NewReno
2419 * and SACK.
2420 *
2421 * Imagine, that's all! Forget about all this shamanism about CWND inflation
2422 * deflation etc. CWND is real congestion window, never inflated, changes
2423 * only according to classic VJ rules.
2424 *
2425 * Really tricky (and requiring careful tuning) part of algorithm
2426 * is hidden in functions tcp_time_to_recover() and tcp_xmit_retransmit_queue().
2427 * The first determines the moment _when_ we should reduce CWND and,
2428 * hence, slow down forward transmission. In fact, it determines the moment
2429 * when we decide that hole is caused by loss, rather than by a reorder.
2430 *
2431 * tcp_xmit_retransmit_queue() decides, _what_ we should retransmit to fill
2432 * holes, caused by lost packets.
2433 *
2434 * And the most logically complicated part of algorithm is undo
2435 * heuristics. We detect false retransmits due to both too early
2436 * fast retransmit (reordering) and underestimated RTO, analyzing
2437 * timestamps and D-SACKs. When we detect that some segments were
2438 * retransmitted by mistake and CWND reduction was wrong, we undo
2439 * window reduction and abort recovery phase. This logic is hidden
2440 * inside several functions named tcp_try_undo_<something>.
2441 */
2442
2443/* This function decides, when we should leave Disordered state
2444 * and enter Recovery phase, reducing congestion window.
2445 *
2446 * Main question: may we further continue forward transmission
2447 * with the same cwnd?
2448 */
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002449static int tcp_time_to_recover(struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002450{
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002451 struct tcp_sock *tp = tcp_sk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002452 __u32 packets_out;
2453
Ryousei Takano564262c2007-10-25 23:03:52 -07002454 /* Do not perform any recovery during F-RTO algorithm */
Ilpo Järvinen52c63f12007-02-21 23:06:52 -08002455 if (tp->frto_counter)
2456 return 0;
2457
Linus Torvalds1da177e2005-04-16 15:20:36 -07002458 /* Trick#1: The loss is proven. */
2459 if (tp->lost_out)
2460 return 1;
2461
2462 /* Not-A-Trick#2 : Classic rule... */
Andreas Petlundea84e552009-10-27 03:27:21 +00002463 if (tcp_dupack_heuristics(tp) > tp->reordering)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002464 return 1;
2465
2466 /* Trick#3 : when we use RFC2988 timer restart, fast
2467 * retransmit can be triggered by timeout of queue head.
2468 */
Ilpo Järvinen85cc3912007-11-15 19:39:31 -08002469 if (tcp_is_fack(tp) && tcp_head_timedout(sk))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002470 return 1;
2471
2472 /* Trick#4: It is still not OK... But will it be useful to delay
2473 * recovery more?
2474 */
2475 packets_out = tp->packets_out;
2476 if (packets_out <= tp->reordering &&
2477 tp->sacked_out >= max_t(__u32, packets_out/2, sysctl_tcp_reordering) &&
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002478 !tcp_may_send_now(sk)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002479 /* We have nothing to send. This connection is limited
2480 * either by receiver window or by application.
2481 */
2482 return 1;
2483 }
2484
Andreas Petlund7e380172010-02-18 04:48:19 +00002485 /* If a thin stream is detected, retransmit after first
2486 * received dupack. Employ only if SACK is supported in order
2487 * to avoid possible corner-case series of spurious retransmissions
2488 * Use only if there are no unsent data.
2489 */
2490 if ((tp->thin_dupack || sysctl_tcp_thin_dupack) &&
2491 tcp_stream_is_thin(tp) && tcp_dupack_heuristics(tp) > 1 &&
2492 tcp_is_sack(tp) && !tcp_send_head(sk))
2493 return 1;
2494
Linus Torvalds1da177e2005-04-16 15:20:36 -07002495 return 0;
2496}
2497
Ilpo Järvinen7363a5b2009-02-28 04:44:33 +00002498/* New heuristics: it is possible only after we switched to restart timer
2499 * each time when something is ACKed. Hence, we can detect timed out packets
2500 * during fast retransmit without falling to slow start.
2501 *
2502 * Usefulness of this as is very questionable, since we should know which of
2503 * the segments is the next to timeout which is relatively expensive to find
2504 * in general case unless we add some data structure just for that. The
2505 * current approach certainly won't find the right one too often and when it
2506 * finally does find _something_ it usually marks large part of the window
2507 * right away (because a retransmission with a larger timestamp blocks the
2508 * loop from advancing). -ij
2509 */
2510static void tcp_timeout_skbs(struct sock *sk)
2511{
2512 struct tcp_sock *tp = tcp_sk(sk);
2513 struct sk_buff *skb;
2514
2515 if (!tcp_is_fack(tp) || !tcp_head_timedout(sk))
2516 return;
2517
2518 skb = tp->scoreboard_skb_hint;
2519 if (tp->scoreboard_skb_hint == NULL)
2520 skb = tcp_write_queue_head(sk);
2521
2522 tcp_for_write_queue_from(skb, sk) {
2523 if (skb == tcp_send_head(sk))
2524 break;
2525 if (!tcp_skb_timedout(sk, skb))
2526 break;
2527
2528 tcp_skb_mark_lost(tp, skb);
2529 }
2530
2531 tp->scoreboard_skb_hint = skb;
2532
2533 tcp_verify_left_out(tp);
2534}
2535
Yuchung Cheng974c1232012-01-19 14:42:21 +00002536/* Detect loss in event "A" above by marking head of queue up as lost.
2537 * For FACK or non-SACK(Reno) senders, the first "packets" number of segments
2538 * are considered lost. For RFC3517 SACK, a segment is considered lost if it
2539 * has at least tp->reordering SACKed seqments above it; "packets" refers to
2540 * the maximum SACKed segments to pass before reaching this limit.
Ilpo Järvinen85cc3912007-11-15 19:39:31 -08002541 */
Ilpo Järvinen1fdb9362010-10-14 01:42:30 +00002542static void tcp_mark_head_lost(struct sock *sk, int packets, int mark_head)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002543{
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002544 struct tcp_sock *tp = tcp_sk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002545 struct sk_buff *skb;
Ilpo Järvinenc137f3d2008-04-07 22:32:38 -07002546 int cnt, oldcnt;
2547 int err;
2548 unsigned int mss;
Yuchung Cheng974c1232012-01-19 14:42:21 +00002549 /* Use SACK to deduce losses of new sequences sent during recovery */
2550 const u32 loss_high = tcp_is_sack(tp) ? tp->snd_nxt : tp->high_seq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002551
Ilpo Järvinen547b7922008-07-25 21:43:18 -07002552 WARN_ON(packets > tp->packets_out);
Stephen Hemminger6a438bb2005-11-10 17:14:59 -08002553 if (tp->lost_skb_hint) {
2554 skb = tp->lost_skb_hint;
2555 cnt = tp->lost_cnt_hint;
Ilpo Järvinen1fdb9362010-10-14 01:42:30 +00002556 /* Head already handled? */
2557 if (mark_head && skb != tcp_write_queue_head(sk))
2558 return;
Stephen Hemminger6a438bb2005-11-10 17:14:59 -08002559 } else {
David S. Millerfe067e82007-03-07 12:12:44 -08002560 skb = tcp_write_queue_head(sk);
Stephen Hemminger6a438bb2005-11-10 17:14:59 -08002561 cnt = 0;
2562 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002563
David S. Millerfe067e82007-03-07 12:12:44 -08002564 tcp_for_write_queue_from(skb, sk) {
2565 if (skb == tcp_send_head(sk))
2566 break;
Stephen Hemminger6a438bb2005-11-10 17:14:59 -08002567 /* TODO: do this better */
2568 /* this is not the most efficient way to do this... */
2569 tp->lost_skb_hint = skb;
2570 tp->lost_cnt_hint = cnt;
Ilpo Järvinen85cc3912007-11-15 19:39:31 -08002571
Yuchung Cheng974c1232012-01-19 14:42:21 +00002572 if (after(TCP_SKB_CB(skb)->end_seq, loss_high))
Ilpo Järvinenc137f3d2008-04-07 22:32:38 -07002573 break;
2574
2575 oldcnt = cnt;
Ilpo Järvinenad1984e2008-01-30 20:06:02 -08002576 if (tcp_is_fack(tp) || tcp_is_reno(tp) ||
Ilpo Järvinen85cc3912007-11-15 19:39:31 -08002577 (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
2578 cnt += tcp_skb_pcount(skb);
2579
Ilpo Järvinenc137f3d2008-04-07 22:32:38 -07002580 if (cnt > packets) {
Yuchung Chengb3de7552010-09-24 13:22:06 +00002581 if ((tcp_is_sack(tp) && !tcp_is_fack(tp)) ||
Neal Cardwellc0638c22012-03-02 21:36:51 +00002582 (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) ||
Yuchung Chengb3de7552010-09-24 13:22:06 +00002583 (oldcnt >= packets))
Ilpo Järvinenc137f3d2008-04-07 22:32:38 -07002584 break;
2585
2586 mss = skb_shinfo(skb)->gso_size;
2587 err = tcp_fragment(sk, skb, (packets - oldcnt) * mss, mss);
2588 if (err < 0)
2589 break;
2590 cnt = packets;
2591 }
2592
Ilpo Järvinen41ea36e2008-09-20 21:19:22 -07002593 tcp_skb_mark_lost(tp, skb);
Ilpo Järvinen1fdb9362010-10-14 01:42:30 +00002594
2595 if (mark_head)
2596 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002597 }
Ilpo Järvinen005903b2007-08-09 14:44:16 +03002598 tcp_verify_left_out(tp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002599}
2600
2601/* Account newly detected lost packet(s) */
2602
Ilpo Järvinen85cc3912007-11-15 19:39:31 -08002603static void tcp_update_scoreboard(struct sock *sk, int fast_rexmit)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002604{
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002605 struct tcp_sock *tp = tcp_sk(sk);
2606
Ilpo Järvinen85cc3912007-11-15 19:39:31 -08002607 if (tcp_is_reno(tp)) {
Ilpo Järvinen1fdb9362010-10-14 01:42:30 +00002608 tcp_mark_head_lost(sk, 1, 1);
Ilpo Järvinen85cc3912007-11-15 19:39:31 -08002609 } else if (tcp_is_fack(tp)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002610 int lost = tp->fackets_out - tp->reordering;
2611 if (lost <= 0)
2612 lost = 1;
Ilpo Järvinen1fdb9362010-10-14 01:42:30 +00002613 tcp_mark_head_lost(sk, lost, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002614 } else {
Ilpo Järvinen85cc3912007-11-15 19:39:31 -08002615 int sacked_upto = tp->sacked_out - tp->reordering;
Ilpo Järvinen1fdb9362010-10-14 01:42:30 +00002616 if (sacked_upto >= 0)
2617 tcp_mark_head_lost(sk, sacked_upto, 0);
2618 else if (fast_rexmit)
2619 tcp_mark_head_lost(sk, 1, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002620 }
2621
Ilpo Järvinen7363a5b2009-02-28 04:44:33 +00002622 tcp_timeout_skbs(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002623}
2624
2625/* CWND moderation, preventing bursts due to too big ACKs
2626 * in dubious situations.
2627 */
2628static inline void tcp_moderate_cwnd(struct tcp_sock *tp)
2629{
2630 tp->snd_cwnd = min(tp->snd_cwnd,
Ilpo Järvinen056834d2007-12-31 14:57:14 -08002631 tcp_packets_in_flight(tp) + tcp_max_burst(tp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002632 tp->snd_cwnd_stamp = tcp_time_stamp;
2633}
2634
Stephen Hemminger72dc5b92006-06-05 17:30:08 -07002635/* Lower bound on congestion window is slow start threshold
2636 * unless congestion avoidance choice decides to overide it.
2637 */
2638static inline u32 tcp_cwnd_min(const struct sock *sk)
2639{
2640 const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops;
2641
2642 return ca_ops->min_cwnd ? ca_ops->min_cwnd(sk) : tcp_sk(sk)->snd_ssthresh;
2643}
2644
Linus Torvalds1da177e2005-04-16 15:20:36 -07002645/* Decrease cwnd each second ack. */
Ilpo Järvinen1e757f92007-07-30 19:48:37 -07002646static void tcp_cwnd_down(struct sock *sk, int flag)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002647{
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002648 struct tcp_sock *tp = tcp_sk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002649 int decr = tp->snd_cwnd_cnt + 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002650
Ilpo Järvinen056834d2007-12-31 14:57:14 -08002651 if ((flag & (FLAG_ANY_PROGRESS | FLAG_DSACKING_ACK)) ||
2652 (tcp_is_reno(tp) && !(flag & FLAG_NOT_DUP))) {
2653 tp->snd_cwnd_cnt = decr & 1;
Ilpo Järvinen1e757f92007-07-30 19:48:37 -07002654 decr >>= 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002655
Ilpo Järvinen1e757f92007-07-30 19:48:37 -07002656 if (decr && tp->snd_cwnd > tcp_cwnd_min(sk))
2657 tp->snd_cwnd -= decr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002658
Ilpo Järvinen056834d2007-12-31 14:57:14 -08002659 tp->snd_cwnd = min(tp->snd_cwnd, tcp_packets_in_flight(tp) + 1);
Ilpo Järvinen1e757f92007-07-30 19:48:37 -07002660 tp->snd_cwnd_stamp = tcp_time_stamp;
2661 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002662}
2663
2664/* Nothing was retransmitted or returned timestamp is less
2665 * than timestamp of the first retransmission.
2666 */
Eric Dumazetcf533ea2011-10-21 05:22:42 -04002667static inline int tcp_packet_delayed(const struct tcp_sock *tp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002668{
2669 return !tp->retrans_stamp ||
2670 (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
Arnd Hannemannd7ee1472008-04-21 14:46:22 -07002671 before(tp->rx_opt.rcv_tsecr, tp->retrans_stamp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002672}
2673
2674/* Undo procedures. */
2675
2676#if FASTRETRANS_DEBUG > 1
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002677static void DBGUNDO(struct sock *sk, const char *msg)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002678{
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002679 struct tcp_sock *tp = tcp_sk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002680 struct inet_sock *inet = inet_sk(sk);
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002681
YOSHIFUJI Hideaki569508c2008-04-14 04:09:36 -07002682 if (sk->sk_family == AF_INET) {
Harvey Harrison673d57e2008-10-31 00:53:57 -07002683 printk(KERN_DEBUG "Undo %s %pI4/%u c%u l%u ss%u/%u p%u\n",
YOSHIFUJI Hideaki569508c2008-04-14 04:09:36 -07002684 msg,
Joe Perches288fcee2010-05-31 23:48:19 -07002685 &inet->inet_daddr, ntohs(inet->inet_dport),
YOSHIFUJI Hideaki569508c2008-04-14 04:09:36 -07002686 tp->snd_cwnd, tcp_left_out(tp),
2687 tp->snd_ssthresh, tp->prior_ssthresh,
2688 tp->packets_out);
2689 }
Eric Dumazetdfd56b82011-12-10 09:48:31 +00002690#if IS_ENABLED(CONFIG_IPV6)
YOSHIFUJI Hideaki569508c2008-04-14 04:09:36 -07002691 else if (sk->sk_family == AF_INET6) {
2692 struct ipv6_pinfo *np = inet6_sk(sk);
Harvey Harrison5b095d9892008-10-29 12:52:50 -07002693 printk(KERN_DEBUG "Undo %s %pI6/%u c%u l%u ss%u/%u p%u\n",
YOSHIFUJI Hideaki569508c2008-04-14 04:09:36 -07002694 msg,
Joe Perches288fcee2010-05-31 23:48:19 -07002695 &np->daddr, ntohs(inet->inet_dport),
YOSHIFUJI Hideaki569508c2008-04-14 04:09:36 -07002696 tp->snd_cwnd, tcp_left_out(tp),
2697 tp->snd_ssthresh, tp->prior_ssthresh,
2698 tp->packets_out);
2699 }
2700#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07002701}
2702#else
2703#define DBGUNDO(x...) do { } while (0)
2704#endif
2705
David S. Millerf6152732011-03-22 19:37:11 -07002706static void tcp_undo_cwr(struct sock *sk, const bool undo_ssthresh)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002707{
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002708 struct tcp_sock *tp = tcp_sk(sk);
2709
Linus Torvalds1da177e2005-04-16 15:20:36 -07002710 if (tp->prior_ssthresh) {
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002711 const struct inet_connection_sock *icsk = inet_csk(sk);
2712
2713 if (icsk->icsk_ca_ops->undo_cwnd)
2714 tp->snd_cwnd = icsk->icsk_ca_ops->undo_cwnd(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002715 else
Ilpo Järvinen056834d2007-12-31 14:57:14 -08002716 tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh << 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002717
Yuchung Cheng67d41202011-03-14 10:57:03 +00002718 if (undo_ssthresh && tp->prior_ssthresh > tp->snd_ssthresh) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002719 tp->snd_ssthresh = tp->prior_ssthresh;
2720 TCP_ECN_withdraw_cwr(tp);
2721 }
2722 } else {
2723 tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh);
2724 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002725 tp->snd_cwnd_stamp = tcp_time_stamp;
2726}
2727
Eric Dumazetcf533ea2011-10-21 05:22:42 -04002728static inline int tcp_may_undo(const struct tcp_sock *tp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002729{
Ilpo Järvinen056834d2007-12-31 14:57:14 -08002730 return tp->undo_marker && (!tp->undo_retrans || tcp_packet_delayed(tp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002731}
2732
2733/* People celebrate: "We love our President!" */
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002734static int tcp_try_undo_recovery(struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002735{
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002736 struct tcp_sock *tp = tcp_sk(sk);
2737
Linus Torvalds1da177e2005-04-16 15:20:36 -07002738 if (tcp_may_undo(tp)) {
Pavel Emelyanov40b215e2008-07-03 01:05:41 -07002739 int mib_idx;
2740
Linus Torvalds1da177e2005-04-16 15:20:36 -07002741 /* Happy end! We did not retransmit anything
2742 * or our original transmission succeeded.
2743 */
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002744 DBGUNDO(sk, inet_csk(sk)->icsk_ca_state == TCP_CA_Loss ? "loss" : "retrans");
David S. Millerf6152732011-03-22 19:37:11 -07002745 tcp_undo_cwr(sk, true);
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002746 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Loss)
Pavel Emelyanov40b215e2008-07-03 01:05:41 -07002747 mib_idx = LINUX_MIB_TCPLOSSUNDO;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002748 else
Pavel Emelyanov40b215e2008-07-03 01:05:41 -07002749 mib_idx = LINUX_MIB_TCPFULLUNDO;
2750
Pavel Emelyanovde0744a2008-07-16 20:31:16 -07002751 NET_INC_STATS_BH(sock_net(sk), mib_idx);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002752 tp->undo_marker = 0;
2753 }
Ilpo Järvinene60402d2007-08-09 15:14:46 +03002754 if (tp->snd_una == tp->high_seq && tcp_is_reno(tp)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002755 /* Hold old state until something *above* high_seq
2756 * is ACKed. For Reno it is MUST to prevent false
2757 * fast retransmits (RFC2582). SACK TCP is safe. */
2758 tcp_moderate_cwnd(tp);
2759 return 1;
2760 }
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002761 tcp_set_ca_state(sk, TCP_CA_Open);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002762 return 0;
2763}
2764
2765/* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002766static void tcp_try_undo_dsack(struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002767{
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002768 struct tcp_sock *tp = tcp_sk(sk);
2769
Linus Torvalds1da177e2005-04-16 15:20:36 -07002770 if (tp->undo_marker && !tp->undo_retrans) {
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002771 DBGUNDO(sk, "D-SACK");
David S. Millerf6152732011-03-22 19:37:11 -07002772 tcp_undo_cwr(sk, true);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002773 tp->undo_marker = 0;
Pavel Emelyanovde0744a2008-07-16 20:31:16 -07002774 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPDSACKUNDO);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002775 }
2776}
2777
Ilpo Järvinen77722b12009-12-08 20:54:11 -08002778/* We can clear retrans_stamp when there are no retransmissions in the
2779 * window. It would seem that it is trivially available for us in
2780 * tp->retrans_out, however, that kind of assumptions doesn't consider
2781 * what will happen if errors occur when sending retransmission for the
2782 * second time. ...It could the that such segment has only
2783 * TCPCB_EVER_RETRANS set at the present time. It seems that checking
2784 * the head skb is enough except for some reneging corner cases that
2785 * are not worth the effort.
2786 *
2787 * Main reason for all this complexity is the fact that connection dying
2788 * time now depends on the validity of the retrans_stamp, in particular,
2789 * that successive retransmissions of a segment must not advance
2790 * retrans_stamp under any conditions.
2791 */
Eric Dumazetcf533ea2011-10-21 05:22:42 -04002792static int tcp_any_retrans_done(const struct sock *sk)
Ilpo Järvinen77722b12009-12-08 20:54:11 -08002793{
Eric Dumazetcf533ea2011-10-21 05:22:42 -04002794 const struct tcp_sock *tp = tcp_sk(sk);
Ilpo Järvinen77722b12009-12-08 20:54:11 -08002795 struct sk_buff *skb;
2796
2797 if (tp->retrans_out)
2798 return 1;
2799
2800 skb = tcp_write_queue_head(sk);
2801 if (unlikely(skb && TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS))
2802 return 1;
2803
2804 return 0;
2805}
2806
Linus Torvalds1da177e2005-04-16 15:20:36 -07002807/* Undo during fast recovery after partial ACK. */
2808
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002809static int tcp_try_undo_partial(struct sock *sk, int acked)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002810{
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002811 struct tcp_sock *tp = tcp_sk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002812 /* Partial ACK arrived. Force Hoe's retransmit. */
Ilpo Järvinen85cc3912007-11-15 19:39:31 -08002813 int failed = tcp_is_reno(tp) || (tcp_fackets_out(tp) > tp->reordering);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002814
2815 if (tcp_may_undo(tp)) {
2816 /* Plain luck! Hole if filled with delayed
2817 * packet, rather than with a retransmit.
2818 */
Ilpo Järvinen77722b12009-12-08 20:54:11 -08002819 if (!tcp_any_retrans_done(sk))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002820 tp->retrans_stamp = 0;
2821
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002822 tcp_update_reordering(sk, tcp_fackets_out(tp) + acked, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002823
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002824 DBGUNDO(sk, "Hoe");
David S. Millerf6152732011-03-22 19:37:11 -07002825 tcp_undo_cwr(sk, false);
Pavel Emelyanovde0744a2008-07-16 20:31:16 -07002826 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPPARTIALUNDO);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002827
2828 /* So... Do not make Hoe's retransmit yet.
2829 * If the first packet was delayed, the rest
2830 * ones are most probably delayed as well.
2831 */
2832 failed = 0;
2833 }
2834 return failed;
2835}
2836
2837/* Undo during loss recovery after partial ACK. */
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002838static int tcp_try_undo_loss(struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002839{
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002840 struct tcp_sock *tp = tcp_sk(sk);
2841
Linus Torvalds1da177e2005-04-16 15:20:36 -07002842 if (tcp_may_undo(tp)) {
2843 struct sk_buff *skb;
David S. Millerfe067e82007-03-07 12:12:44 -08002844 tcp_for_write_queue(skb, sk) {
2845 if (skb == tcp_send_head(sk))
2846 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002847 TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
2848 }
Stephen Hemminger6a438bb2005-11-10 17:14:59 -08002849
Ilpo Järvinen5af4ec22007-09-20 11:30:48 -07002850 tcp_clear_all_retrans_hints(tp);
Stephen Hemminger6a438bb2005-11-10 17:14:59 -08002851
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002852 DBGUNDO(sk, "partial loss");
Linus Torvalds1da177e2005-04-16 15:20:36 -07002853 tp->lost_out = 0;
David S. Millerf6152732011-03-22 19:37:11 -07002854 tcp_undo_cwr(sk, true);
Pavel Emelyanovde0744a2008-07-16 20:31:16 -07002855 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPLOSSUNDO);
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07002856 inet_csk(sk)->icsk_retransmits = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002857 tp->undo_marker = 0;
Ilpo Järvinene60402d2007-08-09 15:14:46 +03002858 if (tcp_is_sack(tp))
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002859 tcp_set_ca_state(sk, TCP_CA_Open);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002860 return 1;
2861 }
2862 return 0;
2863}
2864
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002865static inline void tcp_complete_cwr(struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002866{
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002867 struct tcp_sock *tp = tcp_sk(sk);
Nandita Dukkipatia262f0c2011-08-21 20:21:57 +00002868
2869 /* Do not moderate cwnd if it's already undone in cwr or recovery. */
2870 if (tp->undo_marker) {
Yuchung Cheng1cebce32012-04-30 06:00:18 +00002871 if (inet_csk(sk)->icsk_ca_state == TCP_CA_CWR) {
Nandita Dukkipatia262f0c2011-08-21 20:21:57 +00002872 tp->snd_cwnd = min(tp->snd_cwnd, tp->snd_ssthresh);
Yuchung Cheng1cebce32012-04-30 06:00:18 +00002873 tp->snd_cwnd_stamp = tcp_time_stamp;
2874 } else if (tp->snd_ssthresh < TCP_INFINITE_SSTHRESH) {
2875 /* PRR algorithm. */
Nandita Dukkipatia262f0c2011-08-21 20:21:57 +00002876 tp->snd_cwnd = tp->snd_ssthresh;
Yuchung Cheng1cebce32012-04-30 06:00:18 +00002877 tp->snd_cwnd_stamp = tcp_time_stamp;
2878 }
Yuchung Cheng67d41202011-03-14 10:57:03 +00002879 }
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002880 tcp_ca_event(sk, CA_EVENT_COMPLETE_CWR);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002881}
2882
Ilpo Järvinen8aca6cb2008-06-04 11:34:22 -07002883static void tcp_try_keep_open(struct sock *sk)
2884{
2885 struct tcp_sock *tp = tcp_sk(sk);
2886 int state = TCP_CA_Open;
2887
Neal Cardwellf6982042011-11-16 08:58:04 +00002888 if (tcp_left_out(tp) || tcp_any_retrans_done(sk))
Ilpo Järvinen8aca6cb2008-06-04 11:34:22 -07002889 state = TCP_CA_Disorder;
2890
2891 if (inet_csk(sk)->icsk_ca_state != state) {
2892 tcp_set_ca_state(sk, state);
2893 tp->high_seq = tp->snd_nxt;
2894 }
2895}
2896
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002897static void tcp_try_to_open(struct sock *sk, int flag)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002898{
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002899 struct tcp_sock *tp = tcp_sk(sk);
2900
Ilpo Järvinen86426c22007-08-09 14:45:17 +03002901 tcp_verify_left_out(tp);
2902
Ilpo Järvinen77722b12009-12-08 20:54:11 -08002903 if (!tp->frto_counter && !tcp_any_retrans_done(sk))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002904 tp->retrans_stamp = 0;
2905
Ilpo Järvinen056834d2007-12-31 14:57:14 -08002906 if (flag & FLAG_ECE)
Ilpo Järvinen3cfe3ba2007-02-27 10:09:49 -08002907 tcp_enter_cwr(sk, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002908
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002909 if (inet_csk(sk)->icsk_ca_state != TCP_CA_CWR) {
Ilpo Järvinen8aca6cb2008-06-04 11:34:22 -07002910 tcp_try_keep_open(sk);
Neal Cardwell8cd6d612011-11-16 08:58:05 +00002911 if (inet_csk(sk)->icsk_ca_state != TCP_CA_Open)
2912 tcp_moderate_cwnd(tp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002913 } else {
Ilpo Järvinen1e757f92007-07-30 19:48:37 -07002914 tcp_cwnd_down(sk, flag);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002915 }
2916}
2917
John Heffner5d424d52006-03-20 17:53:41 -08002918static void tcp_mtup_probe_failed(struct sock *sk)
2919{
2920 struct inet_connection_sock *icsk = inet_csk(sk);
2921
2922 icsk->icsk_mtup.search_high = icsk->icsk_mtup.probe_size - 1;
2923 icsk->icsk_mtup.probe_size = 0;
2924}
2925
Ilpo Järvinen72211e92009-03-14 14:23:04 +00002926static void tcp_mtup_probe_success(struct sock *sk)
John Heffner5d424d52006-03-20 17:53:41 -08002927{
2928 struct tcp_sock *tp = tcp_sk(sk);
2929 struct inet_connection_sock *icsk = inet_csk(sk);
2930
2931 /* FIXME: breaks with very large cwnd */
2932 tp->prior_ssthresh = tcp_current_ssthresh(sk);
2933 tp->snd_cwnd = tp->snd_cwnd *
2934 tcp_mss_to_mtu(sk, tp->mss_cache) /
2935 icsk->icsk_mtup.probe_size;
2936 tp->snd_cwnd_cnt = 0;
2937 tp->snd_cwnd_stamp = tcp_time_stamp;
John Heffner9c6d5e52010-10-06 21:18:02 -07002938 tp->snd_ssthresh = tcp_current_ssthresh(sk);
John Heffner5d424d52006-03-20 17:53:41 -08002939
2940 icsk->icsk_mtup.search_low = icsk->icsk_mtup.probe_size;
2941 icsk->icsk_mtup.probe_size = 0;
2942 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
2943}
2944
Ilpo Järvinene1aa6802008-11-24 21:11:55 -08002945/* Do a simple retransmit without using the backoff mechanisms in
2946 * tcp_timer. This is used for path mtu discovery.
2947 * The socket is already locked here.
2948 */
2949void tcp_simple_retransmit(struct sock *sk)
2950{
2951 const struct inet_connection_sock *icsk = inet_csk(sk);
2952 struct tcp_sock *tp = tcp_sk(sk);
2953 struct sk_buff *skb;
Ilpo Järvinen0c54b852009-03-14 14:23:05 +00002954 unsigned int mss = tcp_current_mss(sk);
Ilpo Järvinene1aa6802008-11-24 21:11:55 -08002955 u32 prior_lost = tp->lost_out;
2956
2957 tcp_for_write_queue(skb, sk) {
2958 if (skb == tcp_send_head(sk))
2959 break;
Ilpo Järvinen775ffab2008-12-05 22:41:26 -08002960 if (tcp_skb_seglen(skb) > mss &&
Ilpo Järvinene1aa6802008-11-24 21:11:55 -08002961 !(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) {
2962 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
2963 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
2964 tp->retrans_out -= tcp_skb_pcount(skb);
2965 }
2966 tcp_skb_mark_lost_uncond_verify(tp, skb);
2967 }
2968 }
2969
2970 tcp_clear_retrans_hints_partial(tp);
2971
2972 if (prior_lost == tp->lost_out)
2973 return;
2974
2975 if (tcp_is_reno(tp))
2976 tcp_limit_reno_sacked(tp);
2977
2978 tcp_verify_left_out(tp);
2979
2980 /* Don't muck with the congestion window here.
2981 * Reason is that we do not increase amount of _data_
2982 * in network, but units changed and effective
2983 * cwnd/ssthresh really reduced now.
2984 */
2985 if (icsk->icsk_ca_state != TCP_CA_Loss) {
2986 tp->high_seq = tp->snd_nxt;
2987 tp->snd_ssthresh = tcp_current_ssthresh(sk);
2988 tp->prior_ssthresh = 0;
2989 tp->undo_marker = 0;
2990 tcp_set_ca_state(sk, TCP_CA_Loss);
2991 }
2992 tcp_xmit_retransmit_queue(sk);
2993}
Eric Dumazet4bc2f182010-07-09 21:22:10 +00002994EXPORT_SYMBOL(tcp_simple_retransmit);
Ilpo Järvinene1aa6802008-11-24 21:11:55 -08002995
Nandita Dukkipatia262f0c2011-08-21 20:21:57 +00002996/* This function implements the PRR algorithm, specifcally the PRR-SSRB
2997 * (proportional rate reduction with slow start reduction bound) as described in
2998 * http://www.ietf.org/id/draft-mathis-tcpm-proportional-rate-reduction-01.txt.
2999 * It computes the number of packets to send (sndcnt) based on packets newly
3000 * delivered:
3001 * 1) If the packets in flight is larger than ssthresh, PRR spreads the
3002 * cwnd reductions across a full RTT.
3003 * 2) If packets in flight is lower than ssthresh (such as due to excess
3004 * losses and/or application stalls), do not perform any further cwnd
3005 * reductions, but instead slow start up to ssthresh.
3006 */
3007static void tcp_update_cwnd_in_recovery(struct sock *sk, int newly_acked_sacked,
3008 int fast_rexmit, int flag)
3009{
3010 struct tcp_sock *tp = tcp_sk(sk);
3011 int sndcnt = 0;
3012 int delta = tp->snd_ssthresh - tcp_packets_in_flight(tp);
3013
3014 if (tcp_packets_in_flight(tp) > tp->snd_ssthresh) {
3015 u64 dividend = (u64)tp->snd_ssthresh * tp->prr_delivered +
3016 tp->prior_cwnd - 1;
3017 sndcnt = div_u64(dividend, tp->prior_cwnd) - tp->prr_out;
3018 } else {
3019 sndcnt = min_t(int, delta,
3020 max_t(int, tp->prr_delivered - tp->prr_out,
3021 newly_acked_sacked) + 1);
3022 }
3023
3024 sndcnt = max(sndcnt, (fast_rexmit ? 1 : 0));
3025 tp->snd_cwnd = tcp_packets_in_flight(tp) + sndcnt;
3026}
3027
Linus Torvalds1da177e2005-04-16 15:20:36 -07003028/* Process an event, which can update packets-in-flight not trivially.
3029 * Main goal of this function is to calculate new estimate for left_out,
3030 * taking into account both packets sitting in receiver's buffer and
3031 * packets lost by network.
3032 *
3033 * Besides that it does CWND reduction, when packet loss is detected
3034 * and changes state of machine.
3035 *
3036 * It does _not_ decide what to send, it is made in function
3037 * tcp_xmit_retransmit_queue().
3038 */
Nandita Dukkipatia262f0c2011-08-21 20:21:57 +00003039static void tcp_fastretrans_alert(struct sock *sk, int pkts_acked,
Neal Cardwell7d2b55f2011-11-16 08:58:01 +00003040 int newly_acked_sacked, bool is_dupack,
3041 int flag)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003042{
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03003043 struct inet_connection_sock *icsk = inet_csk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003044 struct tcp_sock *tp = tcp_sk(sk);
Ilpo Järvinen056834d2007-12-31 14:57:14 -08003045 int do_lost = is_dupack || ((flag & FLAG_DATA_SACKED) &&
Ilpo Järvinen85cc3912007-11-15 19:39:31 -08003046 (tcp_fackets_out(tp) > tp->reordering));
Pavel Emelyanov40b215e2008-07-03 01:05:41 -07003047 int fast_rexmit = 0, mib_idx;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003048
Ilpo Järvinen3ccd3132007-12-31 04:43:32 -08003049 if (WARN_ON(!tp->packets_out && tp->sacked_out))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003050 tp->sacked_out = 0;
Ilpo Järvinen91fed7a2007-10-09 01:24:15 -07003051 if (WARN_ON(!tp->sacked_out && tp->fackets_out))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003052 tp->fackets_out = 0;
3053
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09003054 /* Now state machine starts.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003055 * A. ECE, hence prohibit cwnd undoing, the reduction is required. */
Ilpo Järvinen056834d2007-12-31 14:57:14 -08003056 if (flag & FLAG_ECE)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003057 tp->prior_ssthresh = 0;
3058
3059 /* B. In all the states check for reneging SACKs. */
Ilpo Järvinencadbd032007-12-31 04:49:21 -08003060 if (tcp_check_sack_reneging(sk, flag))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003061 return;
3062
Yuchung Cheng974c1232012-01-19 14:42:21 +00003063 /* C. Check consistency of the current state. */
Ilpo Järvinen005903b2007-08-09 14:44:16 +03003064 tcp_verify_left_out(tp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003065
Yuchung Cheng974c1232012-01-19 14:42:21 +00003066 /* D. Check state exit conditions. State can be terminated
Linus Torvalds1da177e2005-04-16 15:20:36 -07003067 * when high_seq is ACKed. */
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03003068 if (icsk->icsk_ca_state == TCP_CA_Open) {
Ilpo Järvinen547b7922008-07-25 21:43:18 -07003069 WARN_ON(tp->retrans_out != 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003070 tp->retrans_stamp = 0;
3071 } else if (!before(tp->snd_una, tp->high_seq)) {
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03003072 switch (icsk->icsk_ca_state) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003073 case TCP_CA_Loss:
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03003074 icsk->icsk_retransmits = 0;
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07003075 if (tcp_try_undo_recovery(sk))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003076 return;
3077 break;
3078
3079 case TCP_CA_CWR:
3080 /* CWR is to be held something *above* high_seq
3081 * is ACKed for CWR bit to reach receiver. */
3082 if (tp->snd_una != tp->high_seq) {
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03003083 tcp_complete_cwr(sk);
3084 tcp_set_ca_state(sk, TCP_CA_Open);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003085 }
3086 break;
3087
Linus Torvalds1da177e2005-04-16 15:20:36 -07003088 case TCP_CA_Recovery:
Ilpo Järvinene60402d2007-08-09 15:14:46 +03003089 if (tcp_is_reno(tp))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003090 tcp_reset_reno_sack(tp);
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07003091 if (tcp_try_undo_recovery(sk))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003092 return;
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03003093 tcp_complete_cwr(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003094 break;
3095 }
3096 }
3097
Yuchung Cheng974c1232012-01-19 14:42:21 +00003098 /* E. Process state. */
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03003099 switch (icsk->icsk_ca_state) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003100 case TCP_CA_Recovery:
Ilpo Järvinen2e605292007-08-02 19:46:58 -07003101 if (!(flag & FLAG_SND_UNA_ADVANCED)) {
Ilpo Järvinene60402d2007-08-09 15:14:46 +03003102 if (tcp_is_reno(tp) && is_dupack)
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03003103 tcp_add_reno_sack(sk);
Ilpo Järvinen1b6d4272007-08-09 14:53:36 +03003104 } else
3105 do_lost = tcp_try_undo_partial(sk, pkts_acked);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003106 break;
3107 case TCP_CA_Loss:
Ilpo Järvinen056834d2007-12-31 14:57:14 -08003108 if (flag & FLAG_DATA_ACKED)
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03003109 icsk->icsk_retransmits = 0;
Ilpo Järvinen882beba2008-04-07 22:33:07 -07003110 if (tcp_is_reno(tp) && flag & FLAG_SND_UNA_ADVANCED)
3111 tcp_reset_reno_sack(tp);
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07003112 if (!tcp_try_undo_loss(sk)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003113 tcp_moderate_cwnd(tp);
3114 tcp_xmit_retransmit_queue(sk);
3115 return;
3116 }
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03003117 if (icsk->icsk_ca_state != TCP_CA_Open)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003118 return;
3119 /* Loss is undone; fall through to processing in Open state. */
3120 default:
Ilpo Järvinene60402d2007-08-09 15:14:46 +03003121 if (tcp_is_reno(tp)) {
Ilpo Järvinen2e605292007-08-02 19:46:58 -07003122 if (flag & FLAG_SND_UNA_ADVANCED)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003123 tcp_reset_reno_sack(tp);
3124 if (is_dupack)
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03003125 tcp_add_reno_sack(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003126 }
3127
Neal Cardwellf6982042011-11-16 08:58:04 +00003128 if (icsk->icsk_ca_state <= TCP_CA_Disorder)
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07003129 tcp_try_undo_dsack(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003130
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07003131 if (!tcp_time_to_recover(sk)) {
3132 tcp_try_to_open(sk, flag);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003133 return;
3134 }
3135
John Heffner5d424d52006-03-20 17:53:41 -08003136 /* MTU probe failure: don't reduce cwnd */
3137 if (icsk->icsk_ca_state < TCP_CA_CWR &&
3138 icsk->icsk_mtup.probe_size &&
John Heffner0e7b1362006-03-20 21:32:58 -08003139 tp->snd_una == tp->mtu_probe.probe_seq_start) {
John Heffner5d424d52006-03-20 17:53:41 -08003140 tcp_mtup_probe_failed(sk);
3141 /* Restores the reduction we did in tcp_mtup_probe() */
3142 tp->snd_cwnd++;
3143 tcp_simple_retransmit(sk);
3144 return;
3145 }
3146
Linus Torvalds1da177e2005-04-16 15:20:36 -07003147 /* Otherwise enter Recovery state */
3148
Ilpo Järvinene60402d2007-08-09 15:14:46 +03003149 if (tcp_is_reno(tp))
Pavel Emelyanov40b215e2008-07-03 01:05:41 -07003150 mib_idx = LINUX_MIB_TCPRENORECOVERY;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003151 else
Pavel Emelyanov40b215e2008-07-03 01:05:41 -07003152 mib_idx = LINUX_MIB_TCPSACKRECOVERY;
3153
Pavel Emelyanovde0744a2008-07-16 20:31:16 -07003154 NET_INC_STATS_BH(sock_net(sk), mib_idx);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003155
3156 tp->high_seq = tp->snd_nxt;
3157 tp->prior_ssthresh = 0;
3158 tp->undo_marker = tp->snd_una;
3159 tp->undo_retrans = tp->retrans_out;
3160
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03003161 if (icsk->icsk_ca_state < TCP_CA_CWR) {
Ilpo Järvinen056834d2007-12-31 14:57:14 -08003162 if (!(flag & FLAG_ECE))
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03003163 tp->prior_ssthresh = tcp_current_ssthresh(sk);
3164 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003165 TCP_ECN_queue_cwr(tp);
3166 }
3167
Stephen Hemminger9772efb2005-11-10 17:09:53 -08003168 tp->bytes_acked = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003169 tp->snd_cwnd_cnt = 0;
Nandita Dukkipatia262f0c2011-08-21 20:21:57 +00003170 tp->prior_cwnd = tp->snd_cwnd;
3171 tp->prr_delivered = 0;
3172 tp->prr_out = 0;
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03003173 tcp_set_ca_state(sk, TCP_CA_Recovery);
Ilpo Järvinen85cc3912007-11-15 19:39:31 -08003174 fast_rexmit = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003175 }
3176
Ilpo Järvinen85cc3912007-11-15 19:39:31 -08003177 if (do_lost || (tcp_is_fack(tp) && tcp_head_timedout(sk)))
3178 tcp_update_scoreboard(sk, fast_rexmit);
Nandita Dukkipatia262f0c2011-08-21 20:21:57 +00003179 tp->prr_delivered += newly_acked_sacked;
3180 tcp_update_cwnd_in_recovery(sk, newly_acked_sacked, fast_rexmit, flag);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003181 tcp_xmit_retransmit_queue(sk);
3182}
3183
Jerry Chu9ad7c042011-06-08 11:08:38 +00003184void tcp_valid_rtt_meas(struct sock *sk, u32 seq_rtt)
Ilpo Järvinen41834b72008-12-05 22:43:26 -08003185{
3186 tcp_rtt_estimator(sk, seq_rtt);
3187 tcp_set_rto(sk);
3188 inet_csk(sk)->icsk_backoff = 0;
3189}
Jerry Chu9ad7c042011-06-08 11:08:38 +00003190EXPORT_SYMBOL(tcp_valid_rtt_meas);
Ilpo Järvinen41834b72008-12-05 22:43:26 -08003191
Linus Torvalds1da177e2005-04-16 15:20:36 -07003192/* Read draft-ietf-tcplw-high-performance before mucking
Stephen Hemmingercaa20d9a2005-11-10 17:13:47 -08003193 * with this code. (Supersedes RFC1323)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003194 */
Stephen Hemminger2d2abba2005-11-10 16:56:12 -08003195static void tcp_ack_saw_tstamp(struct sock *sk, int flag)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003196{
Linus Torvalds1da177e2005-04-16 15:20:36 -07003197 /* RTTM Rule: A TSecr value received in a segment is used to
3198 * update the averaged RTT measurement only if the segment
3199 * acknowledges some new data, i.e., only if it advances the
3200 * left edge of the send window.
3201 *
3202 * See draft-ietf-tcplw-high-performance-00, section 3.3.
3203 * 1998/04/10 Andrey V. Savochkin <saw@msu.ru>
3204 *
3205 * Changed: reset backoff as soon as we see the first valid sample.
Stephen Hemmingercaa20d9a2005-11-10 17:13:47 -08003206 * If we do not, we get strongly overestimated rto. With timestamps
Linus Torvalds1da177e2005-04-16 15:20:36 -07003207 * samples are accepted even from very old segments: f.e., when rtt=1
3208 * increases to 8, we retransmit 5 times and after 8 seconds delayed
3209 * answer arrives rto becomes 120 seconds! If at least one of segments
3210 * in window is lost... Voila. --ANK (010210)
3211 */
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07003212 struct tcp_sock *tp = tcp_sk(sk);
Ilpo Järvinen41834b72008-12-05 22:43:26 -08003213
3214 tcp_valid_rtt_meas(sk, tcp_time_stamp - tp->rx_opt.rcv_tsecr);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003215}
3216
Stephen Hemminger2d2abba2005-11-10 16:56:12 -08003217static void tcp_ack_no_tstamp(struct sock *sk, u32 seq_rtt, int flag)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003218{
3219 /* We don't have a timestamp. Can only use
3220 * packets that are not retransmitted to determine
3221 * rtt estimates. Also, we must not reset the
3222 * backoff for rto until we get a non-retransmitted
3223 * packet. This allows us to deal with a situation
3224 * where the network delay has increased suddenly.
3225 * I.e. Karn's algorithm. (SIGCOMM '87, p5.)
3226 */
3227
3228 if (flag & FLAG_RETRANS_DATA_ACKED)
3229 return;
3230
Ilpo Järvinen41834b72008-12-05 22:43:26 -08003231 tcp_valid_rtt_meas(sk, seq_rtt);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003232}
3233
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07003234static inline void tcp_ack_update_rtt(struct sock *sk, const int flag,
Stephen Hemminger2d2abba2005-11-10 16:56:12 -08003235 const s32 seq_rtt)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003236{
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07003237 const struct tcp_sock *tp = tcp_sk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003238 /* Note that peer MAY send zero echo. In this case it is ignored. (rfc1323) */
3239 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr)
Stephen Hemminger2d2abba2005-11-10 16:56:12 -08003240 tcp_ack_saw_tstamp(sk, flag);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003241 else if (seq_rtt >= 0)
Stephen Hemminger2d2abba2005-11-10 16:56:12 -08003242 tcp_ack_no_tstamp(sk, seq_rtt, flag);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003243}
3244
Ilpo Järvinenc3a05c62007-12-02 00:47:59 +02003245static void tcp_cong_avoid(struct sock *sk, u32 ack, u32 in_flight)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003246{
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03003247 const struct inet_connection_sock *icsk = inet_csk(sk);
Ilpo Järvinenc3a05c62007-12-02 00:47:59 +02003248 icsk->icsk_ca_ops->cong_avoid(sk, ack, in_flight);
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03003249 tcp_sk(sk)->snd_cwnd_stamp = tcp_time_stamp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003250}
3251
Linus Torvalds1da177e2005-04-16 15:20:36 -07003252/* Restart timer after forward progress on connection.
3253 * RFC2988 recommends to restart timer to now+rto.
3254 */
Ilpo Järvinen6728e7d2007-08-24 22:53:26 -07003255static void tcp_rearm_rto(struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003256{
Eric Dumazetcf533ea2011-10-21 05:22:42 -04003257 const struct tcp_sock *tp = tcp_sk(sk);
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07003258
Linus Torvalds1da177e2005-04-16 15:20:36 -07003259 if (!tp->packets_out) {
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07003260 inet_csk_clear_xmit_timer(sk, ICSK_TIME_RETRANS);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003261 } else {
Ilpo Järvinen056834d2007-12-31 14:57:14 -08003262 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
3263 inet_csk(sk)->icsk_rto, TCP_RTO_MAX);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003264 }
3265}
3266
Ilpo Järvinen7c46a032007-09-20 11:33:43 -07003267/* If we get here, the whole TSO packet has not been acked. */
Ilpo Järvinen13fcf852007-10-09 01:28:45 -07003268static u32 tcp_tso_acked(struct sock *sk, struct sk_buff *skb)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003269{
3270 struct tcp_sock *tp = tcp_sk(sk);
Ilpo Järvinen7c46a032007-09-20 11:33:43 -07003271 u32 packets_acked;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003272
Ilpo Järvinen7c46a032007-09-20 11:33:43 -07003273 BUG_ON(!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003274
3275 packets_acked = tcp_skb_pcount(skb);
Ilpo Järvinen7c46a032007-09-20 11:33:43 -07003276 if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003277 return 0;
3278 packets_acked -= tcp_skb_pcount(skb);
3279
3280 if (packets_acked) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003281 BUG_ON(tcp_skb_pcount(skb) == 0);
Ilpo Järvinen7c46a032007-09-20 11:33:43 -07003282 BUG_ON(!before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003283 }
3284
Ilpo Järvinen13fcf852007-10-09 01:28:45 -07003285 return packets_acked;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003286}
3287
Ilpo Järvinen7c46a032007-09-20 11:33:43 -07003288/* Remove acknowledged frames from the retransmission queue. If our packet
3289 * is before the ack sequence we can discard it as it's confirmed to have
3290 * arrived at the other end.
3291 */
Ilpo Järvinen33f5f572008-10-07 14:43:06 -07003292static int tcp_clean_rtx_queue(struct sock *sk, int prior_fackets,
3293 u32 prior_snd_una)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003294{
3295 struct tcp_sock *tp = tcp_sk(sk);
Stephen Hemminger2d2abba2005-11-10 16:56:12 -08003296 const struct inet_connection_sock *icsk = inet_csk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003297 struct sk_buff *skb;
Ilpo Järvinen7c46a032007-09-20 11:33:43 -07003298 u32 now = tcp_time_stamp;
Ilpo Järvinen13fcf852007-10-09 01:28:45 -07003299 int fully_acked = 1;
Ilpo Järvinen7c46a032007-09-20 11:33:43 -07003300 int flag = 0;
Ilpo Järvinen72018832007-12-30 04:37:55 -08003301 u32 pkts_acked = 0;
Ilpo Järvinenc7caf8d2007-11-10 21:22:18 -08003302 u32 reord = tp->packets_out;
Ilpo Järvinen90638a02008-09-20 21:25:52 -07003303 u32 prior_sacked = tp->sacked_out;
Ilpo Järvinen7c46a032007-09-20 11:33:43 -07003304 s32 seq_rtt = -1;
Gavin McCullagh2072c222007-12-29 19:11:21 -08003305 s32 ca_seq_rtt = -1;
Ilpo Järvinenb9ce2042007-06-15 15:08:43 -07003306 ktime_t last_ackt = net_invalid_timestamp();
Linus Torvalds1da177e2005-04-16 15:20:36 -07003307
Ilpo Järvinen7c46a032007-09-20 11:33:43 -07003308 while ((skb = tcp_write_queue_head(sk)) && skb != tcp_send_head(sk)) {
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09003309 struct tcp_skb_cb *scb = TCP_SKB_CB(skb);
Ilpo Järvinen72018832007-12-30 04:37:55 -08003310 u32 acked_pcount;
Ilpo Järvinen7c46a032007-09-20 11:33:43 -07003311 u8 sacked = scb->sacked;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003312
Gavin McCullagh2072c222007-12-29 19:11:21 -08003313 /* Determine how many packets and what bytes were acked, tso and else */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003314 if (after(scb->end_seq, tp->snd_una)) {
Ilpo Järvinen13fcf852007-10-09 01:28:45 -07003315 if (tcp_skb_pcount(skb) == 1 ||
3316 !after(tp->snd_una, scb->seq))
3317 break;
3318
Ilpo Järvinen72018832007-12-30 04:37:55 -08003319 acked_pcount = tcp_tso_acked(sk, skb);
3320 if (!acked_pcount)
Ilpo Järvinen13fcf852007-10-09 01:28:45 -07003321 break;
3322
3323 fully_acked = 0;
Ilpo Järvinen13fcf852007-10-09 01:28:45 -07003324 } else {
Ilpo Järvinen72018832007-12-30 04:37:55 -08003325 acked_pcount = tcp_skb_pcount(skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003326 }
3327
Ilpo Järvinen89d478f2007-12-30 04:35:27 -08003328 if (sacked & TCPCB_RETRANS) {
3329 if (sacked & TCPCB_SACKED_RETRANS)
Ilpo Järvinen72018832007-12-30 04:37:55 -08003330 tp->retrans_out -= acked_pcount;
Ilpo Järvinen89d478f2007-12-30 04:35:27 -08003331 flag |= FLAG_RETRANS_DATA_ACKED;
3332 ca_seq_rtt = -1;
3333 seq_rtt = -1;
Ilpo Järvinen056834d2007-12-31 14:57:14 -08003334 if ((flag & FLAG_DATA_ACKED) || (acked_pcount > 1))
Ilpo Järvinen89d478f2007-12-30 04:35:27 -08003335 flag |= FLAG_NONHEAD_RETRANS_ACKED;
Ilpo Järvinenc7caf8d2007-11-10 21:22:18 -08003336 } else {
Gavin McCullagh2072c222007-12-29 19:11:21 -08003337 ca_seq_rtt = now - scb->when;
3338 last_ackt = skb->tstamp;
Ilpo Järvinenc7caf8d2007-11-10 21:22:18 -08003339 if (seq_rtt < 0) {
Gavin McCullagh2072c222007-12-29 19:11:21 -08003340 seq_rtt = ca_seq_rtt;
Ilpo Järvinenc7caf8d2007-11-10 21:22:18 -08003341 }
Ilpo Järvinen89d478f2007-12-30 04:35:27 -08003342 if (!(sacked & TCPCB_SACKED_ACKED))
Ilpo Järvinen72018832007-12-30 04:37:55 -08003343 reord = min(pkts_acked, reord);
Stephen Hemminger2d2abba2005-11-10 16:56:12 -08003344 }
Ilpo Järvinen89d478f2007-12-30 04:35:27 -08003345
3346 if (sacked & TCPCB_SACKED_ACKED)
Ilpo Järvinen72018832007-12-30 04:37:55 -08003347 tp->sacked_out -= acked_pcount;
Ilpo Järvinen89d478f2007-12-30 04:35:27 -08003348 if (sacked & TCPCB_LOST)
Ilpo Järvinen72018832007-12-30 04:37:55 -08003349 tp->lost_out -= acked_pcount;
Ilpo Järvinen89d478f2007-12-30 04:35:27 -08003350
Ilpo Järvinen72018832007-12-30 04:37:55 -08003351 tp->packets_out -= acked_pcount;
3352 pkts_acked += acked_pcount;
Ilpo Järvinen13fcf852007-10-09 01:28:45 -07003353
Ilpo Järvinen009a2e32007-09-20 11:34:38 -07003354 /* Initial outgoing SYN's get put onto the write_queue
3355 * just like anything else we transmit. It is not
3356 * true data, and if we misinform our callers that
3357 * this ACK acks real data, we will erroneously exit
3358 * connection startup slow start one packet too
3359 * quickly. This is severely frowned upon behavior.
3360 */
Eric Dumazet4de075e2011-09-27 13:25:05 -04003361 if (!(scb->tcp_flags & TCPHDR_SYN)) {
Ilpo Järvinen009a2e32007-09-20 11:34:38 -07003362 flag |= FLAG_DATA_ACKED;
3363 } else {
3364 flag |= FLAG_SYN_ACKED;
3365 tp->retrans_stamp = 0;
3366 }
3367
Ilpo Järvinen13fcf852007-10-09 01:28:45 -07003368 if (!fully_acked)
3369 break;
3370
David S. Millerfe067e82007-03-07 12:12:44 -08003371 tcp_unlink_write_queue(skb, sk);
Hideo Aoki3ab224b2007-12-31 00:11:19 -08003372 sk_wmem_free_skb(sk, skb);
Ilpo Järvinen90638a02008-09-20 21:25:52 -07003373 tp->scoreboard_skb_hint = NULL;
Ilpo Järvinenef9da472008-09-20 21:25:15 -07003374 if (skb == tp->retransmit_skb_hint)
3375 tp->retransmit_skb_hint = NULL;
Ilpo Järvinen90638a02008-09-20 21:25:52 -07003376 if (skb == tp->lost_skb_hint)
3377 tp->lost_skb_hint = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003378 }
3379
Ilpo Järvinen33f5f572008-10-07 14:43:06 -07003380 if (likely(between(tp->snd_up, prior_snd_una, tp->snd_una)))
3381 tp->snd_up = tp->snd_una;
3382
Ilpo Järvinencadbd032007-12-31 04:49:21 -08003383 if (skb && (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
3384 flag |= FLAG_SACK_RENEGING;
3385
Ilpo Järvinen7c46a032007-09-20 11:33:43 -07003386 if (flag & FLAG_ACKED) {
Stephen Hemminger164891a2007-04-23 22:26:16 -07003387 const struct tcp_congestion_ops *ca_ops
3388 = inet_csk(sk)->icsk_ca_ops;
3389
Ilpo Järvinen72211e92009-03-14 14:23:04 +00003390 if (unlikely(icsk->icsk_mtup.probe_size &&
3391 !after(tp->mtu_probe.probe_seq_end, tp->snd_una))) {
3392 tcp_mtup_probe_success(sk);
3393 }
3394
Ilpo Järvinen7c46a032007-09-20 11:33:43 -07003395 tcp_ack_update_rtt(sk, flag, seq_rtt);
Ilpo Järvinen6728e7d2007-08-24 22:53:26 -07003396 tcp_rearm_rto(sk);
Stephen Hemminger317a76f2005-06-23 12:19:55 -07003397
Ilpo Järvinenc7caf8d2007-11-10 21:22:18 -08003398 if (tcp_is_reno(tp)) {
3399 tcp_remove_reno_sacks(sk, pkts_acked);
3400 } else {
Ilpo Järvinen59a08cb2009-02-28 04:44:28 +00003401 int delta;
3402
Ilpo Järvinenc7caf8d2007-11-10 21:22:18 -08003403 /* Non-retransmitted hole got filled? That's reordering */
3404 if (reord < prior_fackets)
3405 tcp_update_reordering(sk, tp->fackets_out - reord, 0);
Ilpo Järvinen90638a02008-09-20 21:25:52 -07003406
Ilpo Järvinen59a08cb2009-02-28 04:44:28 +00003407 delta = tcp_is_fack(tp) ? pkts_acked :
3408 prior_sacked - tp->sacked_out;
3409 tp->lost_cnt_hint -= min(tp->lost_cnt_hint, delta);
Ilpo Järvinenc7caf8d2007-11-10 21:22:18 -08003410 }
3411
Ilpo Järvinen91fed7a2007-10-09 01:24:15 -07003412 tp->fackets_out -= min(pkts_acked, tp->fackets_out);
Ilpo Järvinen68f83532007-11-15 19:50:37 -08003413
Stephen Hemminger30cfd0b2007-07-25 23:49:34 -07003414 if (ca_ops->pkts_acked) {
3415 s32 rtt_us = -1;
Ilpo Järvinenb9ce2042007-06-15 15:08:43 -07003416
Stephen Hemminger30cfd0b2007-07-25 23:49:34 -07003417 /* Is the ACK triggering packet unambiguous? */
Ilpo Järvinen7c46a032007-09-20 11:33:43 -07003418 if (!(flag & FLAG_RETRANS_DATA_ACKED)) {
Stephen Hemminger30cfd0b2007-07-25 23:49:34 -07003419 /* High resolution needed and available? */
3420 if (ca_ops->flags & TCP_CONG_RTT_STAMP &&
3421 !ktime_equal(last_ackt,
3422 net_invalid_timestamp()))
3423 rtt_us = ktime_us_delta(ktime_get_real(),
3424 last_ackt);
stephen hemmingerfebf0812011-03-14 07:52:12 +00003425 else if (ca_seq_rtt >= 0)
Gavin McCullagh2072c222007-12-29 19:11:21 -08003426 rtt_us = jiffies_to_usecs(ca_seq_rtt);
Stephen Hemminger30cfd0b2007-07-25 23:49:34 -07003427 }
3428
3429 ca_ops->pkts_acked(sk, pkts_acked, rtt_us);
3430 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003431 }
3432
3433#if FASTRETRANS_DEBUG > 0
Ilpo Järvinen547b7922008-07-25 21:43:18 -07003434 WARN_ON((int)tp->sacked_out < 0);
3435 WARN_ON((int)tp->lost_out < 0);
3436 WARN_ON((int)tp->retrans_out < 0);
Ilpo Järvinene60402d2007-08-09 15:14:46 +03003437 if (!tp->packets_out && tcp_is_sack(tp)) {
Stephen Hemmingercfcabdc2007-10-09 01:59:42 -07003438 icsk = inet_csk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003439 if (tp->lost_out) {
3440 printk(KERN_DEBUG "Leak l=%u %d\n",
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03003441 tp->lost_out, icsk->icsk_ca_state);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003442 tp->lost_out = 0;
3443 }
3444 if (tp->sacked_out) {
3445 printk(KERN_DEBUG "Leak s=%u %d\n",
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03003446 tp->sacked_out, icsk->icsk_ca_state);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003447 tp->sacked_out = 0;
3448 }
3449 if (tp->retrans_out) {
3450 printk(KERN_DEBUG "Leak r=%u %d\n",
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03003451 tp->retrans_out, icsk->icsk_ca_state);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003452 tp->retrans_out = 0;
3453 }
3454 }
3455#endif
Ilpo Järvinen7c46a032007-09-20 11:33:43 -07003456 return flag;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003457}
3458
3459static void tcp_ack_probe(struct sock *sk)
3460{
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07003461 const struct tcp_sock *tp = tcp_sk(sk);
3462 struct inet_connection_sock *icsk = inet_csk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003463
3464 /* Was it a usable window open? */
3465
Ilpo Järvinen90840de2007-12-31 04:48:41 -08003466 if (!after(TCP_SKB_CB(tcp_send_head(sk))->end_seq, tcp_wnd_end(tp))) {
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07003467 icsk->icsk_backoff = 0;
3468 inet_csk_clear_xmit_timer(sk, ICSK_TIME_PROBE0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003469 /* Socket must be waked up by subsequent tcp_data_snd_check().
3470 * This function is not for random using!
3471 */
3472 } else {
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07003473 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
Arnaldo Carvalho de Melo3f421ba2005-08-09 20:11:08 -07003474 min(icsk->icsk_rto << icsk->icsk_backoff, TCP_RTO_MAX),
3475 TCP_RTO_MAX);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003476 }
3477}
3478
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03003479static inline int tcp_ack_is_dubious(const struct sock *sk, const int flag)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003480{
Eric Dumazeta02cec22010-09-22 20:43:57 +00003481 return !(flag & FLAG_NOT_DUP) || (flag & FLAG_CA_ALERT) ||
3482 inet_csk(sk)->icsk_ca_state != TCP_CA_Open;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003483}
3484
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03003485static inline int tcp_may_raise_cwnd(const struct sock *sk, const int flag)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003486{
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03003487 const struct tcp_sock *tp = tcp_sk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003488 return (!(flag & FLAG_ECE) || tp->snd_cwnd < tp->snd_ssthresh) &&
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03003489 !((1 << inet_csk(sk)->icsk_ca_state) & (TCPF_CA_Recovery | TCPF_CA_CWR));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003490}
3491
3492/* Check that window update is acceptable.
3493 * The function assumes that snd_una<=ack<=snd_next.
3494 */
Ilpo Järvinen056834d2007-12-31 14:57:14 -08003495static inline int tcp_may_update_window(const struct tcp_sock *tp,
3496 const u32 ack, const u32 ack_seq,
3497 const u32 nwin)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003498{
Eric Dumazeta02cec22010-09-22 20:43:57 +00003499 return after(ack, tp->snd_una) ||
Linus Torvalds1da177e2005-04-16 15:20:36 -07003500 after(ack_seq, tp->snd_wl1) ||
Eric Dumazeta02cec22010-09-22 20:43:57 +00003501 (ack_seq == tp->snd_wl1 && nwin > tp->snd_wnd);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003502}
3503
3504/* Update our send window.
3505 *
3506 * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2
3507 * and in FreeBSD. NetBSD's one is even worse.) is wrong.
3508 */
Eric Dumazetcf533ea2011-10-21 05:22:42 -04003509static int tcp_ack_update_window(struct sock *sk, const struct sk_buff *skb, u32 ack,
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07003510 u32 ack_seq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003511{
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07003512 struct tcp_sock *tp = tcp_sk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003513 int flag = 0;
Arnaldo Carvalho de Meloaa8223c2007-04-10 21:04:22 -07003514 u32 nwin = ntohs(tcp_hdr(skb)->window);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003515
Arnaldo Carvalho de Meloaa8223c2007-04-10 21:04:22 -07003516 if (likely(!tcp_hdr(skb)->syn))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003517 nwin <<= tp->rx_opt.snd_wscale;
3518
3519 if (tcp_may_update_window(tp, ack, ack_seq, nwin)) {
3520 flag |= FLAG_WIN_UPDATE;
Hantzis Fotisee7537b2009-03-02 22:42:02 -08003521 tcp_update_wl(tp, ack_seq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003522
3523 if (tp->snd_wnd != nwin) {
3524 tp->snd_wnd = nwin;
3525
3526 /* Note, it is the only place, where
3527 * fast path is recovered for sending TCP.
3528 */
Herbert Xu2ad41062005-10-27 18:47:46 +10003529 tp->pred_flags = 0;
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07003530 tcp_fast_path_check(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003531
3532 if (nwin > tp->max_window) {
3533 tp->max_window = nwin;
Arnaldo Carvalho de Melod83d8462005-12-13 23:26:10 -08003534 tcp_sync_mss(sk, inet_csk(sk)->icsk_pmtu_cookie);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003535 }
3536 }
3537 }
3538
3539 tp->snd_una = ack;
3540
3541 return flag;
3542}
3543
Ilpo Järvinen9ead9a12007-02-21 22:56:19 -08003544/* A very conservative spurious RTO response algorithm: reduce cwnd and
3545 * continue in congestion avoidance.
3546 */
3547static void tcp_conservative_spur_to_response(struct tcp_sock *tp)
3548{
3549 tp->snd_cwnd = min(tp->snd_cwnd, tp->snd_ssthresh);
Ilpo Järvinenaa8b6a72007-02-21 23:06:03 -08003550 tp->snd_cwnd_cnt = 0;
Ilpo Järvinen16e90682007-10-11 17:32:31 -07003551 tp->bytes_acked = 0;
Ilpo Järvinen46323652007-05-19 13:56:23 -07003552 TCP_ECN_queue_cwr(tp);
Ilpo Järvinen9ead9a12007-02-21 22:56:19 -08003553 tcp_moderate_cwnd(tp);
3554}
3555
Ilpo Järvinen3cfe3ba2007-02-27 10:09:49 -08003556/* A conservative spurious RTO response algorithm: reduce cwnd using
3557 * rate halving and continue in congestion avoidance.
3558 */
3559static void tcp_ratehalving_spur_to_response(struct sock *sk)
3560{
Ilpo Järvinen3cfe3ba2007-02-27 10:09:49 -08003561 tcp_enter_cwr(sk, 0);
Ilpo Järvinen3cfe3ba2007-02-27 10:09:49 -08003562}
3563
Ilpo Järvinene317f6f2007-03-02 13:34:19 -08003564static void tcp_undo_spur_to_response(struct sock *sk, int flag)
Ilpo Järvinen3cfe3ba2007-02-27 10:09:49 -08003565{
Ilpo Järvinen056834d2007-12-31 14:57:14 -08003566 if (flag & FLAG_ECE)
Ilpo Järvinene317f6f2007-03-02 13:34:19 -08003567 tcp_ratehalving_spur_to_response(sk);
3568 else
David S. Millerf6152732011-03-22 19:37:11 -07003569 tcp_undo_cwr(sk, true);
Ilpo Järvinen3cfe3ba2007-02-27 10:09:49 -08003570}
3571
Ilpo Järvinen30935cf2007-02-21 23:01:36 -08003572/* F-RTO spurious RTO detection algorithm (RFC4138)
3573 *
Ilpo Järvinen6408d202007-02-21 23:04:11 -08003574 * F-RTO affects during two new ACKs following RTO (well, almost, see inline
3575 * comments). State (ACK number) is kept in frto_counter. When ACK advances
3576 * window (but not to or beyond highest sequence sent before RTO):
Ilpo Järvinen30935cf2007-02-21 23:01:36 -08003577 * On First ACK, send two new segments out.
3578 * On Second ACK, RTO was likely spurious. Do spurious response (response
3579 * algorithm is not part of the F-RTO detection algorithm
3580 * given in RFC4138 but can be selected separately).
3581 * Otherwise (basically on duplicate ACK), RTO was (likely) caused by a loss
Ilpo Järvinend551e452007-04-30 00:42:20 -07003582 * and TCP falls back to conventional RTO recovery. F-RTO allows overriding
3583 * of Nagle, this is done using frto_counter states 2 and 3, when a new data
3584 * segment of any size sent during F-RTO, state 2 is upgraded to 3.
Ilpo Järvinen30935cf2007-02-21 23:01:36 -08003585 *
3586 * Rationale: if the RTO was spurious, new ACKs should arrive from the
3587 * original window even after we transmit two new data segments.
3588 *
Ilpo Järvinen4dc26652007-02-21 23:16:11 -08003589 * SACK version:
3590 * on first step, wait until first cumulative ACK arrives, then move to
3591 * the second step. In second step, the next ACK decides.
3592 *
Ilpo Järvinen30935cf2007-02-21 23:01:36 -08003593 * F-RTO is implemented (mainly) in four functions:
3594 * - tcp_use_frto() is used to determine if TCP is can use F-RTO
3595 * - tcp_enter_frto() prepares TCP state on RTO if F-RTO is used, it is
3596 * called when tcp_use_frto() showed green light
3597 * - tcp_process_frto() handles incoming ACKs during F-RTO algorithm
3598 * - tcp_enter_frto_loss() is called if there is not enough evidence
3599 * to prove that the RTO is indeed spurious. It transfers the control
3600 * from F-RTO to the conventional RTO recovery
3601 */
Ilpo Järvinen2e605292007-08-02 19:46:58 -07003602static int tcp_process_frto(struct sock *sk, int flag)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003603{
3604 struct tcp_sock *tp = tcp_sk(sk);
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09003605
Ilpo Järvinen005903b2007-08-09 14:44:16 +03003606 tcp_verify_left_out(tp);
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09003607
Ilpo Järvinen7487c482007-02-21 23:02:30 -08003608 /* Duplicate the behavior from Loss state (fastretrans_alert) */
Ilpo Järvinen056834d2007-12-31 14:57:14 -08003609 if (flag & FLAG_DATA_ACKED)
Ilpo Järvinen7487c482007-02-21 23:02:30 -08003610 inet_csk(sk)->icsk_retransmits = 0;
3611
Ilpo Järvinen009a2e32007-09-20 11:34:38 -07003612 if ((flag & FLAG_NONHEAD_RETRANS_ACKED) ||
3613 ((tp->frto_counter >= 2) && (flag & FLAG_RETRANS_DATA_ACKED)))
3614 tp->undo_marker = 0;
3615
Ilpo Järvinen95c49222007-02-21 23:05:18 -08003616 if (!before(tp->snd_una, tp->frto_highmark)) {
Ilpo Järvinend551e452007-04-30 00:42:20 -07003617 tcp_enter_frto_loss(sk, (tp->frto_counter == 1 ? 2 : 3), flag);
Ilpo Järvinen7c9a4a52007-02-21 23:08:34 -08003618 return 1;
Ilpo Järvinen95c49222007-02-21 23:05:18 -08003619 }
3620
Ilpo Järvinen62ab2222008-05-08 01:09:11 -07003621 if (!tcp_is_sackfrto(tp)) {
Ilpo Järvinen4dc26652007-02-21 23:16:11 -08003622 /* RFC4138 shortcoming in step 2; should also have case c):
3623 * ACK isn't duplicate nor advances window, e.g., opposite dir
3624 * data, winupdate
3625 */
Ilpo Järvinen056834d2007-12-31 14:57:14 -08003626 if (!(flag & FLAG_ANY_PROGRESS) && (flag & FLAG_NOT_DUP))
Ilpo Järvinen4dc26652007-02-21 23:16:11 -08003627 return 1;
Ilpo Järvinen6408d202007-02-21 23:04:11 -08003628
Ilpo Järvinen056834d2007-12-31 14:57:14 -08003629 if (!(flag & FLAG_DATA_ACKED)) {
Ilpo Järvinen4dc26652007-02-21 23:16:11 -08003630 tcp_enter_frto_loss(sk, (tp->frto_counter == 1 ? 0 : 3),
3631 flag);
3632 return 1;
3633 }
3634 } else {
Ilpo Järvinen056834d2007-12-31 14:57:14 -08003635 if (!(flag & FLAG_DATA_ACKED) && (tp->frto_counter == 1)) {
Ilpo Järvinen4dc26652007-02-21 23:16:11 -08003636 /* Prevent sending of new data. */
3637 tp->snd_cwnd = min(tp->snd_cwnd,
3638 tcp_packets_in_flight(tp));
3639 return 1;
3640 }
3641
Ilpo Järvinend551e452007-04-30 00:42:20 -07003642 if ((tp->frto_counter >= 2) &&
Ilpo Järvinen056834d2007-12-31 14:57:14 -08003643 (!(flag & FLAG_FORWARD_PROGRESS) ||
3644 ((flag & FLAG_DATA_SACKED) &&
3645 !(flag & FLAG_ONLY_ORIG_SACKED)))) {
Ilpo Järvinen4dc26652007-02-21 23:16:11 -08003646 /* RFC4138 shortcoming (see comment above) */
Ilpo Järvinen056834d2007-12-31 14:57:14 -08003647 if (!(flag & FLAG_FORWARD_PROGRESS) &&
3648 (flag & FLAG_NOT_DUP))
Ilpo Järvinen4dc26652007-02-21 23:16:11 -08003649 return 1;
3650
3651 tcp_enter_frto_loss(sk, 3, flag);
3652 return 1;
3653 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003654 }
3655
3656 if (tp->frto_counter == 1) {
Ilpo Järvinen3e6f0492007-12-05 02:20:21 -08003657 /* tcp_may_send_now needs to see updated state */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003658 tp->snd_cwnd = tcp_packets_in_flight(tp) + 2;
Ilpo Järvinen94d0ea72007-02-21 23:07:27 -08003659 tp->frto_counter = 2;
Ilpo Järvinen3e6f0492007-12-05 02:20:21 -08003660
3661 if (!tcp_may_send_now(sk))
3662 tcp_enter_frto_loss(sk, 2, flag);
3663
Ilpo Järvinen7c9a4a52007-02-21 23:08:34 -08003664 return 1;
Ilpo Järvinend551e452007-04-30 00:42:20 -07003665 } else {
Ilpo Järvinen3cfe3ba2007-02-27 10:09:49 -08003666 switch (sysctl_tcp_frto_response) {
3667 case 2:
Ilpo Järvinene317f6f2007-03-02 13:34:19 -08003668 tcp_undo_spur_to_response(sk, flag);
Ilpo Järvinen3cfe3ba2007-02-27 10:09:49 -08003669 break;
3670 case 1:
3671 tcp_conservative_spur_to_response(tp);
3672 break;
3673 default:
3674 tcp_ratehalving_spur_to_response(sk);
3675 break;
Stephen Hemminger3ff50b72007-04-20 17:09:22 -07003676 }
Ilpo Järvinen94d0ea72007-02-21 23:07:27 -08003677 tp->frto_counter = 0;
Ilpo Järvinen009a2e32007-09-20 11:34:38 -07003678 tp->undo_marker = 0;
Pavel Emelyanovde0744a2008-07-16 20:31:16 -07003679 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPSPURIOUSRTOS);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003680 }
Ilpo Järvinen7c9a4a52007-02-21 23:08:34 -08003681 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003682}
3683
Linus Torvalds1da177e2005-04-16 15:20:36 -07003684/* This routine deals with incoming acks, but not outgoing ones. */
Eric Dumazetcf533ea2011-10-21 05:22:42 -04003685static int tcp_ack(struct sock *sk, const struct sk_buff *skb, int flag)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003686{
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03003687 struct inet_connection_sock *icsk = inet_csk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003688 struct tcp_sock *tp = tcp_sk(sk);
3689 u32 prior_snd_una = tp->snd_una;
3690 u32 ack_seq = TCP_SKB_CB(skb)->seq;
3691 u32 ack = TCP_SKB_CB(skb)->ack_seq;
Neal Cardwell7d2b55f2011-11-16 08:58:01 +00003692 bool is_dupack = false;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003693 u32 prior_in_flight;
Ilpo Järvinenc7caf8d2007-11-10 21:22:18 -08003694 u32 prior_fackets;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003695 int prior_packets;
Nandita Dukkipatia262f0c2011-08-21 20:21:57 +00003696 int prior_sacked = tp->sacked_out;
Neal Cardwell7d2b55f2011-11-16 08:58:01 +00003697 int pkts_acked = 0;
Nandita Dukkipatia262f0c2011-08-21 20:21:57 +00003698 int newly_acked_sacked = 0;
Ilpo Järvinen7c9a4a52007-02-21 23:08:34 -08003699 int frto_cwnd = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003700
John Dykstra96e0bf42009-03-22 21:49:57 -07003701 /* If the ack is older than previous acks
Linus Torvalds1da177e2005-04-16 15:20:36 -07003702 * then we can probably ignore it.
3703 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003704 if (before(ack, prior_snd_una))
3705 goto old_ack;
3706
John Dykstra96e0bf42009-03-22 21:49:57 -07003707 /* If the ack includes data we haven't sent yet, discard
3708 * this segment (RFC793 Section 3.9).
3709 */
3710 if (after(ack, tp->snd_nxt))
3711 goto invalid_ack;
3712
Ilpo Järvinen2e605292007-08-02 19:46:58 -07003713 if (after(ack, prior_snd_una))
3714 flag |= FLAG_SND_UNA_ADVANCED;
3715
Daikichi Osuga3fdf3f02006-08-29 02:01:44 -07003716 if (sysctl_tcp_abc) {
3717 if (icsk->icsk_ca_state < TCP_CA_CWR)
3718 tp->bytes_acked += ack - prior_snd_una;
3719 else if (icsk->icsk_ca_state == TCP_CA_Loss)
3720 /* we assume just one segment left network */
Ilpo Järvinen056834d2007-12-31 14:57:14 -08003721 tp->bytes_acked += min(ack - prior_snd_una,
3722 tp->mss_cache);
Daikichi Osuga3fdf3f02006-08-29 02:01:44 -07003723 }
Stephen Hemminger9772efb2005-11-10 17:09:53 -08003724
Ilpo Järvinenc7caf8d2007-11-10 21:22:18 -08003725 prior_fackets = tp->fackets_out;
Ilpo Järvinen52d34082007-12-05 02:21:35 -08003726 prior_in_flight = tcp_packets_in_flight(tp);
Ilpo Järvinenc7caf8d2007-11-10 21:22:18 -08003727
Ilpo Järvinen056834d2007-12-31 14:57:14 -08003728 if (!(flag & FLAG_SLOWPATH) && after(ack, prior_snd_una)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003729 /* Window is constant, pure forward advance.
3730 * No more checks are required.
3731 * Note, we use the fact that SND.UNA>=SND.WL2.
3732 */
Hantzis Fotisee7537b2009-03-02 22:42:02 -08003733 tcp_update_wl(tp, ack_seq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003734 tp->snd_una = ack;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003735 flag |= FLAG_WIN_UPDATE;
3736
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03003737 tcp_ca_event(sk, CA_EVENT_FAST_ACK);
Stephen Hemminger317a76f2005-06-23 12:19:55 -07003738
Pavel Emelyanovde0744a2008-07-16 20:31:16 -07003739 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPHPACKS);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003740 } else {
3741 if (ack_seq != TCP_SKB_CB(skb)->end_seq)
3742 flag |= FLAG_DATA;
3743 else
Pavel Emelyanovde0744a2008-07-16 20:31:16 -07003744 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPPUREACKS);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003745
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07003746 flag |= tcp_ack_update_window(sk, skb, ack, ack_seq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003747
3748 if (TCP_SKB_CB(skb)->sacked)
3749 flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una);
3750
Arnaldo Carvalho de Meloaa8223c2007-04-10 21:04:22 -07003751 if (TCP_ECN_rcv_ecn_echo(tp, tcp_hdr(skb)))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003752 flag |= FLAG_ECE;
3753
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03003754 tcp_ca_event(sk, CA_EVENT_SLOW_ACK);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003755 }
3756
3757 /* We passed data and got it acked, remove any soft error
3758 * log. Something worked...
3759 */
3760 sk->sk_err_soft = 0;
David S. Miller4b53fb62008-07-23 16:38:45 -07003761 icsk->icsk_probes_out = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003762 tp->rcv_tstamp = tcp_time_stamp;
3763 prior_packets = tp->packets_out;
3764 if (!prior_packets)
3765 goto no_queue;
3766
Linus Torvalds1da177e2005-04-16 15:20:36 -07003767 /* See if we can take anything off of the retransmit queue. */
Ilpo Järvinen33f5f572008-10-07 14:43:06 -07003768 flag |= tcp_clean_rtx_queue(sk, prior_fackets, prior_snd_una);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003769
Neal Cardwell7d2b55f2011-11-16 08:58:01 +00003770 pkts_acked = prior_packets - tp->packets_out;
Nandita Dukkipatia262f0c2011-08-21 20:21:57 +00003771 newly_acked_sacked = (prior_packets - prior_sacked) -
3772 (tp->packets_out - tp->sacked_out);
3773
Ilpo Järvinene1cd8f72007-11-14 15:55:09 -08003774 if (tp->frto_counter)
3775 frto_cwnd = tcp_process_frto(sk, flag);
Ilpo Järvinen3de96472007-10-01 15:28:48 -07003776 /* Guarantee sacktag reordering detection against wrap-arounds */
3777 if (before(tp->frto_highmark, tp->snd_una))
3778 tp->frto_highmark = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003779
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03003780 if (tcp_ack_is_dubious(sk, flag)) {
Stephen Hemmingercaa20d9a2005-11-10 17:13:47 -08003781 /* Advance CWND, if state allows this. */
Ilpo Järvinen7c9a4a52007-02-21 23:08:34 -08003782 if ((flag & FLAG_DATA_ACKED) && !frto_cwnd &&
3783 tcp_may_raise_cwnd(sk, flag))
Ilpo Järvinenc3a05c62007-12-02 00:47:59 +02003784 tcp_cong_avoid(sk, ack, prior_in_flight);
Neal Cardwell7d2b55f2011-11-16 08:58:01 +00003785 is_dupack = !(flag & (FLAG_SND_UNA_ADVANCED | FLAG_NOT_DUP));
3786 tcp_fastretrans_alert(sk, pkts_acked, newly_acked_sacked,
3787 is_dupack, flag);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003788 } else {
Ilpo Järvinen7c9a4a52007-02-21 23:08:34 -08003789 if ((flag & FLAG_DATA_ACKED) && !frto_cwnd)
Ilpo Järvinenc3a05c62007-12-02 00:47:59 +02003790 tcp_cong_avoid(sk, ack, prior_in_flight);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003791 }
3792
Ilpo Järvinen056834d2007-12-31 14:57:14 -08003793 if ((flag & FLAG_FORWARD_PROGRESS) || !(flag & FLAG_NOT_DUP))
Eric Dumazetb6c67122010-04-08 23:03:29 +00003794 dst_confirm(__sk_dst_get(sk));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003795
3796 return 1;
3797
3798no_queue:
Neal Cardwell5628adf2011-11-16 08:58:02 +00003799 /* If data was DSACKed, see if we can undo a cwnd reduction. */
3800 if (flag & FLAG_DSACKING_ACK)
3801 tcp_fastretrans_alert(sk, pkts_acked, newly_acked_sacked,
3802 is_dupack, flag);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003803 /* If this ack opens up a zero window, clear backoff. It was
3804 * being used to time the probes, and is probably far higher than
3805 * it needs to be for normal retransmission.
3806 */
David S. Millerfe067e82007-03-07 12:12:44 -08003807 if (tcp_send_head(sk))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003808 tcp_ack_probe(sk);
3809 return 1;
3810
John Dykstra96e0bf42009-03-22 21:49:57 -07003811invalid_ack:
3812 SOCK_DEBUG(sk, "Ack %u after %u:%u\n", ack, tp->snd_una, tp->snd_nxt);
3813 return -1;
3814
Linus Torvalds1da177e2005-04-16 15:20:36 -07003815old_ack:
Neal Cardwelle95ae2f2011-11-16 08:58:03 +00003816 /* If data was SACKed, tag it and see if we should send more data.
3817 * If data was DSACKed, see if we can undo a cwnd reduction.
3818 */
Ilpo Järvinen8aca6cb2008-06-04 11:34:22 -07003819 if (TCP_SKB_CB(skb)->sacked) {
Neal Cardwelle95ae2f2011-11-16 08:58:03 +00003820 flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una);
3821 newly_acked_sacked = tp->sacked_out - prior_sacked;
3822 tcp_fastretrans_alert(sk, pkts_acked, newly_acked_sacked,
3823 is_dupack, flag);
Ilpo Järvinen8aca6cb2008-06-04 11:34:22 -07003824 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003825
John Dykstra96e0bf42009-03-22 21:49:57 -07003826 SOCK_DEBUG(sk, "Ack %u before %u:%u\n", ack, tp->snd_una, tp->snd_nxt);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003827 return 0;
3828}
3829
Linus Torvalds1da177e2005-04-16 15:20:36 -07003830/* Look for tcp options. Normally only called on SYN and SYNACK packets.
3831 * But, this can also be called on packets in the established flow when
3832 * the fast version below fails.
3833 */
Eric Dumazetcf533ea2011-10-21 05:22:42 -04003834void tcp_parse_options(const struct sk_buff *skb, struct tcp_options_received *opt_rx,
3835 const u8 **hvpp, int estab)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003836{
Eric Dumazetcf533ea2011-10-21 05:22:42 -04003837 const unsigned char *ptr;
3838 const struct tcphdr *th = tcp_hdr(skb);
Ilpo Järvinen056834d2007-12-31 14:57:14 -08003839 int length = (th->doff * 4) - sizeof(struct tcphdr);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003840
Eric Dumazetcf533ea2011-10-21 05:22:42 -04003841 ptr = (const unsigned char *)(th + 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003842 opt_rx->saw_tstamp = 0;
3843
Stephen Hemminger2de979b2007-03-08 20:45:19 -08003844 while (length > 0) {
Ilpo Järvinen056834d2007-12-31 14:57:14 -08003845 int opcode = *ptr++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003846 int opsize;
3847
3848 switch (opcode) {
Ilpo Järvinenf038ac82008-01-03 20:36:55 -08003849 case TCPOPT_EOL:
3850 return;
3851 case TCPOPT_NOP: /* Ref: RFC 793 section 3.1 */
3852 length--;
3853 continue;
3854 default:
3855 opsize = *ptr++;
3856 if (opsize < 2) /* "silly options" */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003857 return;
Ilpo Järvinenf038ac82008-01-03 20:36:55 -08003858 if (opsize > length)
3859 return; /* don't parse partial options */
3860 switch (opcode) {
3861 case TCPOPT_MSS:
3862 if (opsize == TCPOLEN_MSS && th->syn && !estab) {
Harvey Harrisond3e2ce32008-05-02 16:26:16 -07003863 u16 in_mss = get_unaligned_be16(ptr);
Ilpo Järvinenf038ac82008-01-03 20:36:55 -08003864 if (in_mss) {
3865 if (opt_rx->user_mss &&
3866 opt_rx->user_mss < in_mss)
3867 in_mss = opt_rx->user_mss;
3868 opt_rx->mss_clamp = in_mss;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003869 }
Stephen Hemminger3ff50b72007-04-20 17:09:22 -07003870 }
Ilpo Järvinenf038ac82008-01-03 20:36:55 -08003871 break;
3872 case TCPOPT_WINDOW:
3873 if (opsize == TCPOLEN_WINDOW && th->syn &&
David S. Millerbb5b7c12009-12-15 20:56:42 -08003874 !estab && sysctl_tcp_window_scaling) {
Ilpo Järvinenf038ac82008-01-03 20:36:55 -08003875 __u8 snd_wscale = *(__u8 *)ptr;
3876 opt_rx->wscale_ok = 1;
3877 if (snd_wscale > 14) {
3878 if (net_ratelimit())
Joe Perches058bd4d2012-03-11 18:36:11 +00003879 pr_info("%s: Illegal window scaling value %d >14 received\n",
3880 __func__,
3881 snd_wscale);
Ilpo Järvinenf038ac82008-01-03 20:36:55 -08003882 snd_wscale = 14;
3883 }
3884 opt_rx->snd_wscale = snd_wscale;
3885 }
3886 break;
3887 case TCPOPT_TIMESTAMP:
3888 if ((opsize == TCPOLEN_TIMESTAMP) &&
3889 ((estab && opt_rx->tstamp_ok) ||
David S. Millerbb5b7c12009-12-15 20:56:42 -08003890 (!estab && sysctl_tcp_timestamps))) {
Ilpo Järvinenf038ac82008-01-03 20:36:55 -08003891 opt_rx->saw_tstamp = 1;
Harvey Harrisond3e2ce32008-05-02 16:26:16 -07003892 opt_rx->rcv_tsval = get_unaligned_be32(ptr);
3893 opt_rx->rcv_tsecr = get_unaligned_be32(ptr + 4);
Ilpo Järvinenf038ac82008-01-03 20:36:55 -08003894 }
3895 break;
3896 case TCPOPT_SACK_PERM:
3897 if (opsize == TCPOLEN_SACK_PERM && th->syn &&
David S. Millerbb5b7c12009-12-15 20:56:42 -08003898 !estab && sysctl_tcp_sack) {
Vijay Subramanianab562222011-12-20 13:23:24 +00003899 opt_rx->sack_ok = TCP_SACK_SEEN;
Ilpo Järvinenf038ac82008-01-03 20:36:55 -08003900 tcp_sack_reset(opt_rx);
3901 }
3902 break;
Stephen Hemminger3ff50b72007-04-20 17:09:22 -07003903
Ilpo Järvinenf038ac82008-01-03 20:36:55 -08003904 case TCPOPT_SACK:
3905 if ((opsize >= (TCPOLEN_SACK_BASE + TCPOLEN_SACK_PERBLOCK)) &&
3906 !((opsize - TCPOLEN_SACK_BASE) % TCPOLEN_SACK_PERBLOCK) &&
3907 opt_rx->sack_ok) {
3908 TCP_SKB_CB(skb)->sacked = (ptr - 2) - (unsigned char *)th;
3909 }
3910 break;
3911#ifdef CONFIG_TCP_MD5SIG
3912 case TCPOPT_MD5SIG:
3913 /*
3914 * The MD5 Hash has already been
3915 * checked (see tcp_v{4,6}_do_rcv()).
3916 */
3917 break;
3918#endif
William Allen Simpson4957faa2009-12-02 18:25:27 +00003919 case TCPOPT_COOKIE:
3920 /* This option is variable length.
3921 */
3922 switch (opsize) {
3923 case TCPOLEN_COOKIE_BASE:
3924 /* not yet implemented */
3925 break;
3926 case TCPOLEN_COOKIE_PAIR:
3927 /* not yet implemented */
3928 break;
3929 case TCPOLEN_COOKIE_MIN+0:
3930 case TCPOLEN_COOKIE_MIN+2:
3931 case TCPOLEN_COOKIE_MIN+4:
3932 case TCPOLEN_COOKIE_MIN+6:
3933 case TCPOLEN_COOKIE_MAX:
3934 /* 16-bit multiple */
3935 opt_rx->cookie_plus = opsize;
3936 *hvpp = ptr;
Joe Perchesccbd6a52010-05-14 10:58:26 +00003937 break;
William Allen Simpson4957faa2009-12-02 18:25:27 +00003938 default:
3939 /* ignore option */
3940 break;
Joe Perchesccbd6a52010-05-14 10:58:26 +00003941 }
William Allen Simpson4957faa2009-12-02 18:25:27 +00003942 break;
Joe Perchesccbd6a52010-05-14 10:58:26 +00003943 }
Ilpo Järvinenf038ac82008-01-03 20:36:55 -08003944
3945 ptr += opsize-2;
3946 length -= opsize;
Stephen Hemminger3ff50b72007-04-20 17:09:22 -07003947 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003948 }
3949}
Eric Dumazet4bc2f182010-07-09 21:22:10 +00003950EXPORT_SYMBOL(tcp_parse_options);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003951
Eric Dumazetcf533ea2011-10-21 05:22:42 -04003952static int tcp_parse_aligned_timestamp(struct tcp_sock *tp, const struct tcphdr *th)
Ilpo Järvinena4356b22008-08-23 05:12:29 -07003953{
Eric Dumazetcf533ea2011-10-21 05:22:42 -04003954 const __be32 *ptr = (const __be32 *)(th + 1);
Ilpo Järvinena4356b22008-08-23 05:12:29 -07003955
3956 if (*ptr == htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16)
3957 | (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP)) {
3958 tp->rx_opt.saw_tstamp = 1;
3959 ++ptr;
3960 tp->rx_opt.rcv_tsval = ntohl(*ptr);
3961 ++ptr;
3962 tp->rx_opt.rcv_tsecr = ntohl(*ptr);
3963 return 1;
3964 }
3965 return 0;
3966}
3967
Linus Torvalds1da177e2005-04-16 15:20:36 -07003968/* Fast parse options. This hopes to only see timestamps.
3969 * If it is wrong it falls back on tcp_parse_options().
3970 */
Eric Dumazetcf533ea2011-10-21 05:22:42 -04003971static int tcp_fast_parse_options(const struct sk_buff *skb,
3972 const struct tcphdr *th,
3973 struct tcp_sock *tp, const u8 **hvpp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003974{
William Allen Simpson4957faa2009-12-02 18:25:27 +00003975 /* In the spirit of fast parsing, compare doff directly to constant
3976 * values. Because equality is used, short doff can be ignored here.
3977 */
3978 if (th->doff == (sizeof(*th) / 4)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003979 tp->rx_opt.saw_tstamp = 0;
3980 return 0;
3981 } else if (tp->rx_opt.tstamp_ok &&
William Allen Simpson4957faa2009-12-02 18:25:27 +00003982 th->doff == ((sizeof(*th) + TCPOLEN_TSTAMP_ALIGNED) / 4)) {
Ilpo Järvinena4356b22008-08-23 05:12:29 -07003983 if (tcp_parse_aligned_timestamp(tp, th))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003984 return 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003985 }
David S. Millerbb5b7c12009-12-15 20:56:42 -08003986 tcp_parse_options(skb, &tp->rx_opt, hvpp, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003987 return 1;
3988}
3989
YOSHIFUJI Hideaki7d5d5522008-04-17 12:29:53 +09003990#ifdef CONFIG_TCP_MD5SIG
3991/*
3992 * Parse MD5 Signature option
3993 */
Eric Dumazetcf533ea2011-10-21 05:22:42 -04003994const u8 *tcp_parse_md5sig_option(const struct tcphdr *th)
YOSHIFUJI Hideaki7d5d5522008-04-17 12:29:53 +09003995{
Eric Dumazetcf533ea2011-10-21 05:22:42 -04003996 int length = (th->doff << 2) - sizeof(*th);
3997 const u8 *ptr = (const u8 *)(th + 1);
YOSHIFUJI Hideaki7d5d5522008-04-17 12:29:53 +09003998
3999 /* If the TCP option is too short, we can short cut */
4000 if (length < TCPOLEN_MD5SIG)
4001 return NULL;
4002
4003 while (length > 0) {
4004 int opcode = *ptr++;
4005 int opsize;
4006
4007 switch(opcode) {
4008 case TCPOPT_EOL:
4009 return NULL;
4010 case TCPOPT_NOP:
4011 length--;
4012 continue;
4013 default:
4014 opsize = *ptr++;
4015 if (opsize < 2 || opsize > length)
4016 return NULL;
4017 if (opcode == TCPOPT_MD5SIG)
Dmitry Popovba78e2d2010-08-07 20:24:28 -07004018 return opsize == TCPOLEN_MD5SIG ? ptr : NULL;
YOSHIFUJI Hideaki7d5d5522008-04-17 12:29:53 +09004019 }
4020 ptr += opsize - 2;
4021 length -= opsize;
4022 }
4023 return NULL;
4024}
Eric Dumazet4bc2f182010-07-09 21:22:10 +00004025EXPORT_SYMBOL(tcp_parse_md5sig_option);
YOSHIFUJI Hideaki7d5d5522008-04-17 12:29:53 +09004026#endif
4027
Linus Torvalds1da177e2005-04-16 15:20:36 -07004028static inline void tcp_store_ts_recent(struct tcp_sock *tp)
4029{
4030 tp->rx_opt.ts_recent = tp->rx_opt.rcv_tsval;
James Morris9d729f72007-03-04 16:12:44 -08004031 tp->rx_opt.ts_recent_stamp = get_seconds();
Linus Torvalds1da177e2005-04-16 15:20:36 -07004032}
4033
4034static inline void tcp_replace_ts_recent(struct tcp_sock *tp, u32 seq)
4035{
4036 if (tp->rx_opt.saw_tstamp && !after(seq, tp->rcv_wup)) {
4037 /* PAWS bug workaround wrt. ACK frames, the PAWS discard
4038 * extra check below makes sure this can only happen
4039 * for pure ACK frames. -DaveM
4040 *
4041 * Not only, also it occurs for expired timestamps.
4042 */
4043
Ilpo Järvinenc887e6d2009-03-14 14:23:03 +00004044 if (tcp_paws_check(&tp->rx_opt, 0))
Linus Torvalds1da177e2005-04-16 15:20:36 -07004045 tcp_store_ts_recent(tp);
4046 }
4047}
4048
4049/* Sorry, PAWS as specified is broken wrt. pure-ACKs -DaveM
4050 *
4051 * It is not fatal. If this ACK does _not_ change critical state (seqs, window)
4052 * it can pass through stack. So, the following predicate verifies that
4053 * this segment is not used for anything but congestion avoidance or
4054 * fast retransmit. Moreover, we even are able to eliminate most of such
4055 * second order effects, if we apply some small "replay" window (~RTO)
4056 * to timestamp space.
4057 *
4058 * All these measures still do not guarantee that we reject wrapped ACKs
4059 * on networks with high bandwidth, when sequence space is recycled fastly,
4060 * but it guarantees that such events will be very rare and do not affect
4061 * connection seriously. This doesn't look nice, but alas, PAWS is really
4062 * buggy extension.
4063 *
4064 * [ Later note. Even worse! It is buggy for segments _with_ data. RFC
4065 * states that events when retransmit arrives after original data are rare.
4066 * It is a blatant lie. VJ forgot about fast retransmit! 8)8) It is
4067 * the biggest problem on large power networks even with minor reordering.
4068 * OK, let's give it small replay window. If peer clock is even 1hz, it is safe
4069 * up to bandwidth of 18Gigabit/sec. 8) ]
4070 */
4071
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07004072static int tcp_disordered_ack(const struct sock *sk, const struct sk_buff *skb)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004073{
Eric Dumazetcf533ea2011-10-21 05:22:42 -04004074 const struct tcp_sock *tp = tcp_sk(sk);
4075 const struct tcphdr *th = tcp_hdr(skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004076 u32 seq = TCP_SKB_CB(skb)->seq;
4077 u32 ack = TCP_SKB_CB(skb)->ack_seq;
4078
4079 return (/* 1. Pure ACK with correct sequence number. */
4080 (th->ack && seq == TCP_SKB_CB(skb)->end_seq && seq == tp->rcv_nxt) &&
4081
4082 /* 2. ... and duplicate ACK. */
4083 ack == tp->snd_una &&
4084
4085 /* 3. ... and does not update window. */
4086 !tcp_may_update_window(tp, ack, seq, ntohs(th->window) << tp->rx_opt.snd_wscale) &&
4087
4088 /* 4. ... and sits in replay window. */
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07004089 (s32)(tp->rx_opt.ts_recent - tp->rx_opt.rcv_tsval) <= (inet_csk(sk)->icsk_rto * 1024) / HZ);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004090}
4091
Ilpo Järvinen056834d2007-12-31 14:57:14 -08004092static inline int tcp_paws_discard(const struct sock *sk,
4093 const struct sk_buff *skb)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004094{
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07004095 const struct tcp_sock *tp = tcp_sk(sk);
Ilpo Järvinenc887e6d2009-03-14 14:23:03 +00004096
4097 return !tcp_paws_check(&tp->rx_opt, TCP_PAWS_WINDOW) &&
4098 !tcp_disordered_ack(sk, skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004099}
4100
4101/* Check segment sequence number for validity.
4102 *
4103 * Segment controls are considered valid, if the segment
4104 * fits to the window after truncation to the window. Acceptability
4105 * of data (and SYN, FIN, of course) is checked separately.
4106 * See tcp_data_queue(), for example.
4107 *
4108 * Also, controls (RST is main one) are accepted using RCV.WUP instead
4109 * of RCV.NXT. Peer still did not advance his SND.UNA when we
4110 * delayed ACK, so that hisSND.UNA<=ourRCV.WUP.
4111 * (borrowed from freebsd)
4112 */
4113
Eric Dumazetcf533ea2011-10-21 05:22:42 -04004114static inline int tcp_sequence(const struct tcp_sock *tp, u32 seq, u32 end_seq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004115{
4116 return !before(end_seq, tp->rcv_wup) &&
4117 !after(seq, tp->rcv_nxt + tcp_receive_window(tp));
4118}
4119
4120/* When we get a reset we do this. */
4121static void tcp_reset(struct sock *sk)
4122{
4123 /* We want the right error as BSD sees it (and indeed as we do). */
4124 switch (sk->sk_state) {
Ilpo Järvinen056834d2007-12-31 14:57:14 -08004125 case TCP_SYN_SENT:
4126 sk->sk_err = ECONNREFUSED;
4127 break;
4128 case TCP_CLOSE_WAIT:
4129 sk->sk_err = EPIPE;
4130 break;
4131 case TCP_CLOSE:
4132 return;
4133 default:
4134 sk->sk_err = ECONNRESET;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004135 }
Tom Marshalla4d25802010-09-20 15:42:05 -07004136 /* This barrier is coupled with smp_rmb() in tcp_poll() */
4137 smp_wmb();
Linus Torvalds1da177e2005-04-16 15:20:36 -07004138
4139 if (!sock_flag(sk, SOCK_DEAD))
4140 sk->sk_error_report(sk);
4141
4142 tcp_done(sk);
4143}
4144
4145/*
4146 * Process the FIN bit. This now behaves as it is supposed to work
4147 * and the FIN takes effect when it is validly part of sequence
4148 * space. Not before when we get holes.
4149 *
4150 * If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT
4151 * (and thence onto LAST-ACK and finally, CLOSE, we never enter
4152 * TIME-WAIT)
4153 *
4154 * If we are in FINWAIT-1, a received FIN indicates simultaneous
4155 * close and we go into CLOSING (and later onto TIME-WAIT)
4156 *
4157 * If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT.
4158 */
Eric Dumazet20c4cb72011-10-20 17:44:03 -04004159static void tcp_fin(struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004160{
4161 struct tcp_sock *tp = tcp_sk(sk);
4162
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07004163 inet_csk_schedule_ack(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004164
4165 sk->sk_shutdown |= RCV_SHUTDOWN;
4166 sock_set_flag(sk, SOCK_DONE);
4167
4168 switch (sk->sk_state) {
Ilpo Järvinen056834d2007-12-31 14:57:14 -08004169 case TCP_SYN_RECV:
4170 case TCP_ESTABLISHED:
4171 /* Move to CLOSE_WAIT */
4172 tcp_set_state(sk, TCP_CLOSE_WAIT);
4173 inet_csk(sk)->icsk_ack.pingpong = 1;
4174 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004175
Ilpo Järvinen056834d2007-12-31 14:57:14 -08004176 case TCP_CLOSE_WAIT:
4177 case TCP_CLOSING:
4178 /* Received a retransmission of the FIN, do
4179 * nothing.
4180 */
4181 break;
4182 case TCP_LAST_ACK:
4183 /* RFC793: Remain in the LAST-ACK state. */
4184 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004185
Ilpo Järvinen056834d2007-12-31 14:57:14 -08004186 case TCP_FIN_WAIT1:
4187 /* This case occurs when a simultaneous close
4188 * happens, we must ack the received FIN and
4189 * enter the CLOSING state.
4190 */
4191 tcp_send_ack(sk);
4192 tcp_set_state(sk, TCP_CLOSING);
4193 break;
4194 case TCP_FIN_WAIT2:
4195 /* Received a FIN -- send ACK and enter TIME_WAIT. */
4196 tcp_send_ack(sk);
4197 tcp_time_wait(sk, TCP_TIME_WAIT, 0);
4198 break;
4199 default:
4200 /* Only TCP_LISTEN and TCP_CLOSE are left, in these
4201 * cases we should never reach this piece of code.
4202 */
Joe Perches058bd4d2012-03-11 18:36:11 +00004203 pr_err("%s: Impossible, sk->sk_state=%d\n",
Harvey Harrison0dc47872008-03-05 20:47:47 -08004204 __func__, sk->sk_state);
Ilpo Järvinen056834d2007-12-31 14:57:14 -08004205 break;
Stephen Hemminger3ff50b72007-04-20 17:09:22 -07004206 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07004207
4208 /* It _is_ possible, that we have something out-of-order _after_ FIN.
4209 * Probably, we should reset in this case. For now drop them.
4210 */
4211 __skb_queue_purge(&tp->out_of_order_queue);
Ilpo Järvinene60402d2007-08-09 15:14:46 +03004212 if (tcp_is_sack(tp))
Linus Torvalds1da177e2005-04-16 15:20:36 -07004213 tcp_sack_reset(&tp->rx_opt);
Hideo Aoki3ab224b2007-12-31 00:11:19 -08004214 sk_mem_reclaim(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004215
4216 if (!sock_flag(sk, SOCK_DEAD)) {
4217 sk->sk_state_change(sk);
4218
4219 /* Do not send POLL_HUP for half duplex close. */
4220 if (sk->sk_shutdown == SHUTDOWN_MASK ||
4221 sk->sk_state == TCP_CLOSE)
Pavel Emelyanov8d8ad9d2007-11-26 20:10:50 +08004222 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_HUP);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004223 else
Pavel Emelyanov8d8ad9d2007-11-26 20:10:50 +08004224 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_IN);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004225 }
4226}
4227
Ilpo Järvinen056834d2007-12-31 14:57:14 -08004228static inline int tcp_sack_extend(struct tcp_sack_block *sp, u32 seq,
4229 u32 end_seq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004230{
4231 if (!after(seq, sp->end_seq) && !after(sp->start_seq, end_seq)) {
4232 if (before(seq, sp->start_seq))
4233 sp->start_seq = seq;
4234 if (after(end_seq, sp->end_seq))
4235 sp->end_seq = end_seq;
4236 return 1;
4237 }
4238 return 0;
4239}
4240
Pavel Emelyanov1ed83462008-07-16 20:29:51 -07004241static void tcp_dsack_set(struct sock *sk, u32 seq, u32 end_seq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004242{
Pavel Emelyanov1ed83462008-07-16 20:29:51 -07004243 struct tcp_sock *tp = tcp_sk(sk);
4244
David S. Millerbb5b7c12009-12-15 20:56:42 -08004245 if (tcp_is_sack(tp) && sysctl_tcp_dsack) {
Pavel Emelyanov40b215e2008-07-03 01:05:41 -07004246 int mib_idx;
4247
Linus Torvalds1da177e2005-04-16 15:20:36 -07004248 if (before(seq, tp->rcv_nxt))
Pavel Emelyanov40b215e2008-07-03 01:05:41 -07004249 mib_idx = LINUX_MIB_TCPDSACKOLDSENT;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004250 else
Pavel Emelyanov40b215e2008-07-03 01:05:41 -07004251 mib_idx = LINUX_MIB_TCPDSACKOFOSENT;
4252
Pavel Emelyanovde0744a2008-07-16 20:31:16 -07004253 NET_INC_STATS_BH(sock_net(sk), mib_idx);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004254
4255 tp->rx_opt.dsack = 1;
4256 tp->duplicate_sack[0].start_seq = seq;
4257 tp->duplicate_sack[0].end_seq = end_seq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004258 }
4259}
4260
Pavel Emelyanov1ed83462008-07-16 20:29:51 -07004261static void tcp_dsack_extend(struct sock *sk, u32 seq, u32 end_seq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004262{
Pavel Emelyanov1ed83462008-07-16 20:29:51 -07004263 struct tcp_sock *tp = tcp_sk(sk);
4264
Linus Torvalds1da177e2005-04-16 15:20:36 -07004265 if (!tp->rx_opt.dsack)
Pavel Emelyanov1ed83462008-07-16 20:29:51 -07004266 tcp_dsack_set(sk, seq, end_seq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004267 else
4268 tcp_sack_extend(tp->duplicate_sack, seq, end_seq);
4269}
4270
Eric Dumazetcf533ea2011-10-21 05:22:42 -04004271static void tcp_send_dupack(struct sock *sk, const struct sk_buff *skb)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004272{
4273 struct tcp_sock *tp = tcp_sk(sk);
4274
4275 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
4276 before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
Pavel Emelyanovde0744a2008-07-16 20:31:16 -07004277 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_DELAYEDACKLOST);
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07004278 tcp_enter_quickack_mode(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004279
David S. Millerbb5b7c12009-12-15 20:56:42 -08004280 if (tcp_is_sack(tp) && sysctl_tcp_dsack) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07004281 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
4282
4283 if (after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt))
4284 end_seq = tp->rcv_nxt;
Pavel Emelyanov1ed83462008-07-16 20:29:51 -07004285 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, end_seq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004286 }
4287 }
4288
4289 tcp_send_ack(sk);
4290}
4291
4292/* These routines update the SACK block as out-of-order packets arrive or
4293 * in-order packets close up the sequence space.
4294 */
4295static void tcp_sack_maybe_coalesce(struct tcp_sock *tp)
4296{
4297 int this_sack;
4298 struct tcp_sack_block *sp = &tp->selective_acks[0];
Ilpo Järvinen056834d2007-12-31 14:57:14 -08004299 struct tcp_sack_block *swalk = sp + 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004300
4301 /* See if the recent change to the first SACK eats into
4302 * or hits the sequence space of other SACK blocks, if so coalesce.
4303 */
Ilpo Järvinen056834d2007-12-31 14:57:14 -08004304 for (this_sack = 1; this_sack < tp->rx_opt.num_sacks;) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07004305 if (tcp_sack_extend(sp, swalk->start_seq, swalk->end_seq)) {
4306 int i;
4307
4308 /* Zap SWALK, by moving every further SACK up by one slot.
4309 * Decrease num_sacks.
4310 */
4311 tp->rx_opt.num_sacks--;
Ilpo Järvinen056834d2007-12-31 14:57:14 -08004312 for (i = this_sack; i < tp->rx_opt.num_sacks; i++)
4313 sp[i] = sp[i + 1];
Linus Torvalds1da177e2005-04-16 15:20:36 -07004314 continue;
4315 }
4316 this_sack++, swalk++;
4317 }
4318}
4319
Linus Torvalds1da177e2005-04-16 15:20:36 -07004320static void tcp_sack_new_ofo_skb(struct sock *sk, u32 seq, u32 end_seq)
4321{
4322 struct tcp_sock *tp = tcp_sk(sk);
4323 struct tcp_sack_block *sp = &tp->selective_acks[0];
4324 int cur_sacks = tp->rx_opt.num_sacks;
4325 int this_sack;
4326
4327 if (!cur_sacks)
4328 goto new_sack;
4329
Ilpo Järvinen056834d2007-12-31 14:57:14 -08004330 for (this_sack = 0; this_sack < cur_sacks; this_sack++, sp++) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07004331 if (tcp_sack_extend(sp, seq, end_seq)) {
4332 /* Rotate this_sack to the first one. */
Ilpo Järvinen056834d2007-12-31 14:57:14 -08004333 for (; this_sack > 0; this_sack--, sp--)
Ilpo Järvinena0bffff2009-03-21 13:36:17 -07004334 swap(*sp, *(sp - 1));
Linus Torvalds1da177e2005-04-16 15:20:36 -07004335 if (cur_sacks > 1)
4336 tcp_sack_maybe_coalesce(tp);
4337 return;
4338 }
4339 }
4340
4341 /* Could not find an adjacent existing SACK, build a new one,
4342 * put it at the front, and shift everyone else down. We
4343 * always know there is at least one SACK present already here.
4344 *
4345 * If the sack array is full, forget about the last one.
4346 */
Adam Langley4389dde2008-07-19 00:07:02 -07004347 if (this_sack >= TCP_NUM_SACKS) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07004348 this_sack--;
4349 tp->rx_opt.num_sacks--;
4350 sp--;
4351 }
Stephen Hemminger2de979b2007-03-08 20:45:19 -08004352 for (; this_sack > 0; this_sack--, sp--)
Ilpo Järvinen056834d2007-12-31 14:57:14 -08004353 *sp = *(sp - 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004354
4355new_sack:
4356 /* Build the new head SACK, and we're done. */
4357 sp->start_seq = seq;
4358 sp->end_seq = end_seq;
4359 tp->rx_opt.num_sacks++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004360}
4361
4362/* RCV.NXT advances, some SACKs should be eaten. */
4363
4364static void tcp_sack_remove(struct tcp_sock *tp)
4365{
4366 struct tcp_sack_block *sp = &tp->selective_acks[0];
4367 int num_sacks = tp->rx_opt.num_sacks;
4368 int this_sack;
4369
4370 /* Empty ofo queue, hence, all the SACKs are eaten. Clear. */
David S. Millerb03efcf2005-07-08 14:57:23 -07004371 if (skb_queue_empty(&tp->out_of_order_queue)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07004372 tp->rx_opt.num_sacks = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004373 return;
4374 }
4375
Ilpo Järvinen056834d2007-12-31 14:57:14 -08004376 for (this_sack = 0; this_sack < num_sacks;) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07004377 /* Check if the start of the sack is covered by RCV.NXT. */
4378 if (!before(tp->rcv_nxt, sp->start_seq)) {
4379 int i;
4380
4381 /* RCV.NXT must cover all the block! */
Ilpo Järvinen547b7922008-07-25 21:43:18 -07004382 WARN_ON(before(tp->rcv_nxt, sp->end_seq));
Linus Torvalds1da177e2005-04-16 15:20:36 -07004383
4384 /* Zap this SACK, by moving forward any other SACKS. */
4385 for (i=this_sack+1; i < num_sacks; i++)
4386 tp->selective_acks[i-1] = tp->selective_acks[i];
4387 num_sacks--;
4388 continue;
4389 }
4390 this_sack++;
4391 sp++;
4392 }
Ilpo Järvinen5861f8e2009-03-14 14:23:01 +00004393 tp->rx_opt.num_sacks = num_sacks;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004394}
4395
4396/* This one checks to see if we can put data from the
4397 * out_of_order queue into the receive_queue.
4398 */
4399static void tcp_ofo_queue(struct sock *sk)
4400{
4401 struct tcp_sock *tp = tcp_sk(sk);
4402 __u32 dsack_high = tp->rcv_nxt;
4403 struct sk_buff *skb;
4404
4405 while ((skb = skb_peek(&tp->out_of_order_queue)) != NULL) {
4406 if (after(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
4407 break;
4408
4409 if (before(TCP_SKB_CB(skb)->seq, dsack_high)) {
4410 __u32 dsack = dsack_high;
4411 if (before(TCP_SKB_CB(skb)->end_seq, dsack_high))
4412 dsack_high = TCP_SKB_CB(skb)->end_seq;
Pavel Emelyanov1ed83462008-07-16 20:29:51 -07004413 tcp_dsack_extend(sk, TCP_SKB_CB(skb)->seq, dsack);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004414 }
4415
4416 if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
Frans Popb1383382010-03-24 07:57:28 +00004417 SOCK_DEBUG(sk, "ofo packet was already received\n");
David S. Miller8728b832005-08-09 19:25:21 -07004418 __skb_unlink(skb, &tp->out_of_order_queue);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004419 __kfree_skb(skb);
4420 continue;
4421 }
4422 SOCK_DEBUG(sk, "ofo requeuing : rcv_next %X seq %X - %X\n",
4423 tp->rcv_nxt, TCP_SKB_CB(skb)->seq,
4424 TCP_SKB_CB(skb)->end_seq);
4425
David S. Miller8728b832005-08-09 19:25:21 -07004426 __skb_unlink(skb, &tp->out_of_order_queue);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004427 __skb_queue_tail(&sk->sk_receive_queue, skb);
4428 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
Arnaldo Carvalho de Meloaa8223c2007-04-10 21:04:22 -07004429 if (tcp_hdr(skb)->fin)
Eric Dumazet20c4cb72011-10-20 17:44:03 -04004430 tcp_fin(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004431 }
4432}
4433
Vitaliy Gusev56f367b2008-04-15 20:26:34 -07004434static int tcp_prune_ofo_queue(struct sock *sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004435static int tcp_prune_queue(struct sock *sk);
4436
Vitaliy Gusevb000cd32008-04-15 00:33:38 -07004437static inline int tcp_try_rmem_schedule(struct sock *sk, unsigned int size)
4438{
4439 if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf ||
4440 !sk_rmem_schedule(sk, size)) {
4441
4442 if (tcp_prune_queue(sk) < 0)
4443 return -1;
4444
4445 if (!sk_rmem_schedule(sk, size)) {
Vitaliy Gusev56f367b2008-04-15 20:26:34 -07004446 if (!tcp_prune_ofo_queue(sk))
4447 return -1;
4448
Vitaliy Gusevb000cd32008-04-15 00:33:38 -07004449 if (!sk_rmem_schedule(sk, size))
4450 return -1;
4451 }
4452 }
4453 return 0;
4454}
4455
Eric Dumazete86b2912012-03-18 11:06:44 +00004456static void tcp_data_queue_ofo(struct sock *sk, struct sk_buff *skb)
4457{
4458 struct tcp_sock *tp = tcp_sk(sk);
4459 struct sk_buff *skb1;
4460 u32 seq, end_seq;
4461
4462 TCP_ECN_check_ce(tp, skb);
4463
4464 if (tcp_try_rmem_schedule(sk, skb->truesize)) {
4465 /* TODO: should increment a counter */
4466 __kfree_skb(skb);
4467 return;
4468 }
4469
4470 /* Disable header prediction. */
4471 tp->pred_flags = 0;
4472 inet_csk_schedule_ack(sk);
4473
4474 SOCK_DEBUG(sk, "out of order segment: rcv_next %X seq %X - %X\n",
4475 tp->rcv_nxt, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
4476
4477 skb1 = skb_peek_tail(&tp->out_of_order_queue);
4478 if (!skb1) {
4479 /* Initial out of order segment, build 1 SACK. */
4480 if (tcp_is_sack(tp)) {
4481 tp->rx_opt.num_sacks = 1;
4482 tp->selective_acks[0].start_seq = TCP_SKB_CB(skb)->seq;
4483 tp->selective_acks[0].end_seq =
4484 TCP_SKB_CB(skb)->end_seq;
4485 }
4486 __skb_queue_head(&tp->out_of_order_queue, skb);
4487 goto end;
4488 }
4489
4490 seq = TCP_SKB_CB(skb)->seq;
4491 end_seq = TCP_SKB_CB(skb)->end_seq;
4492
4493 if (seq == TCP_SKB_CB(skb1)->end_seq) {
Eric Dumazetc8628152012-03-18 11:07:47 +00004494 /* Packets in ofo can stay in queue a long time.
4495 * Better try to coalesce them right now
4496 * to avoid future tcp_collapse_ofo_queue(),
4497 * probably the most expensive function in tcp stack.
4498 */
4499 if (skb->len <= skb_tailroom(skb1) && !tcp_hdr(skb)->fin) {
4500 NET_INC_STATS_BH(sock_net(sk),
4501 LINUX_MIB_TCPRCVCOALESCE);
4502 BUG_ON(skb_copy_bits(skb, 0,
4503 skb_put(skb1, skb->len),
4504 skb->len));
4505 TCP_SKB_CB(skb1)->end_seq = end_seq;
4506 TCP_SKB_CB(skb1)->ack_seq = TCP_SKB_CB(skb)->ack_seq;
4507 __kfree_skb(skb);
4508 skb = NULL;
4509 } else {
4510 __skb_queue_after(&tp->out_of_order_queue, skb1, skb);
4511 }
Eric Dumazete86b2912012-03-18 11:06:44 +00004512
4513 if (!tp->rx_opt.num_sacks ||
4514 tp->selective_acks[0].end_seq != seq)
4515 goto add_sack;
4516
4517 /* Common case: data arrive in order after hole. */
4518 tp->selective_acks[0].end_seq = end_seq;
4519 goto end;
4520 }
4521
4522 /* Find place to insert this segment. */
4523 while (1) {
4524 if (!after(TCP_SKB_CB(skb1)->seq, seq))
4525 break;
4526 if (skb_queue_is_first(&tp->out_of_order_queue, skb1)) {
4527 skb1 = NULL;
4528 break;
4529 }
4530 skb1 = skb_queue_prev(&tp->out_of_order_queue, skb1);
4531 }
4532
4533 /* Do skb overlap to previous one? */
4534 if (skb1 && before(seq, TCP_SKB_CB(skb1)->end_seq)) {
4535 if (!after(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
4536 /* All the bits are present. Drop. */
4537 __kfree_skb(skb);
4538 skb = NULL;
4539 tcp_dsack_set(sk, seq, end_seq);
4540 goto add_sack;
4541 }
4542 if (after(seq, TCP_SKB_CB(skb1)->seq)) {
4543 /* Partial overlap. */
4544 tcp_dsack_set(sk, seq,
4545 TCP_SKB_CB(skb1)->end_seq);
4546 } else {
4547 if (skb_queue_is_first(&tp->out_of_order_queue,
4548 skb1))
4549 skb1 = NULL;
4550 else
4551 skb1 = skb_queue_prev(
4552 &tp->out_of_order_queue,
4553 skb1);
4554 }
4555 }
4556 if (!skb1)
4557 __skb_queue_head(&tp->out_of_order_queue, skb);
4558 else
4559 __skb_queue_after(&tp->out_of_order_queue, skb1, skb);
4560
4561 /* And clean segments covered by new one as whole. */
4562 while (!skb_queue_is_last(&tp->out_of_order_queue, skb)) {
4563 skb1 = skb_queue_next(&tp->out_of_order_queue, skb);
4564
4565 if (!after(end_seq, TCP_SKB_CB(skb1)->seq))
4566 break;
4567 if (before(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
4568 tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq,
4569 end_seq);
4570 break;
4571 }
4572 __skb_unlink(skb1, &tp->out_of_order_queue);
4573 tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq,
4574 TCP_SKB_CB(skb1)->end_seq);
4575 __kfree_skb(skb1);
4576 }
4577
4578add_sack:
4579 if (tcp_is_sack(tp))
4580 tcp_sack_new_ofo_skb(sk, seq, end_seq);
4581end:
4582 if (skb)
4583 skb_set_owner_r(skb, sk);
4584}
4585
4586
Linus Torvalds1da177e2005-04-16 15:20:36 -07004587static void tcp_data_queue(struct sock *sk, struct sk_buff *skb)
4588{
Eric Dumazetcf533ea2011-10-21 05:22:42 -04004589 const struct tcphdr *th = tcp_hdr(skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004590 struct tcp_sock *tp = tcp_sk(sk);
4591 int eaten = -1;
4592
4593 if (TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq)
4594 goto drop;
4595
Eric Dumazetf84af322010-04-28 15:31:51 -07004596 skb_dst_drop(skb);
Ilpo Järvinen056834d2007-12-31 14:57:14 -08004597 __skb_pull(skb, th->doff * 4);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004598
4599 TCP_ECN_accept_cwr(tp, skb);
4600
Ilpo Järvinen5861f8e2009-03-14 14:23:01 +00004601 tp->rx_opt.dsack = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004602
4603 /* Queue data for delivery to the user.
4604 * Packets in sequence go to the receive queue.
4605 * Out of sequence packets to the out_of_order_queue.
4606 */
4607 if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt) {
4608 if (tcp_receive_window(tp) == 0)
4609 goto out_of_window;
4610
4611 /* Ok. In sequence. In window. */
4612 if (tp->ucopy.task == current &&
4613 tp->copied_seq == tp->rcv_nxt && tp->ucopy.len &&
4614 sock_owned_by_user(sk) && !tp->urg_data) {
4615 int chunk = min_t(unsigned int, skb->len,
Ilpo Järvinen056834d2007-12-31 14:57:14 -08004616 tp->ucopy.len);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004617
4618 __set_current_state(TASK_RUNNING);
4619
4620 local_bh_enable();
4621 if (!skb_copy_datagram_iovec(skb, 0, tp->ucopy.iov, chunk)) {
4622 tp->ucopy.len -= chunk;
4623 tp->copied_seq += chunk;
Jerry Chu44f53242011-01-25 13:46:30 -08004624 eaten = (chunk == skb->len);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004625 tcp_rcv_space_adjust(sk);
4626 }
4627 local_bh_disable();
4628 }
4629
4630 if (eaten <= 0) {
4631queue_and_out:
4632 if (eaten < 0 &&
Vitaliy Gusevb000cd32008-04-15 00:33:38 -07004633 tcp_try_rmem_schedule(sk, skb->truesize))
4634 goto drop;
4635
Hideo Aoki3ab224b2007-12-31 00:11:19 -08004636 skb_set_owner_r(skb, sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004637 __skb_queue_tail(&sk->sk_receive_queue, skb);
4638 }
4639 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
Stephen Hemminger2de979b2007-03-08 20:45:19 -08004640 if (skb->len)
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07004641 tcp_event_data_recv(sk, skb);
Stephen Hemminger2de979b2007-03-08 20:45:19 -08004642 if (th->fin)
Eric Dumazet20c4cb72011-10-20 17:44:03 -04004643 tcp_fin(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004644
David S. Millerb03efcf2005-07-08 14:57:23 -07004645 if (!skb_queue_empty(&tp->out_of_order_queue)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07004646 tcp_ofo_queue(sk);
4647
4648 /* RFC2581. 4.2. SHOULD send immediate ACK, when
4649 * gap in queue is filled.
4650 */
David S. Millerb03efcf2005-07-08 14:57:23 -07004651 if (skb_queue_empty(&tp->out_of_order_queue))
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07004652 inet_csk(sk)->icsk_ack.pingpong = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004653 }
4654
4655 if (tp->rx_opt.num_sacks)
4656 tcp_sack_remove(tp);
4657
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07004658 tcp_fast_path_check(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004659
4660 if (eaten > 0)
4661 __kfree_skb(skb);
4662 else if (!sock_flag(sk, SOCK_DEAD))
4663 sk->sk_data_ready(sk, 0);
4664 return;
4665 }
4666
4667 if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
4668 /* A retransmit, 2nd most common case. Force an immediate ack. */
Pavel Emelyanovde0744a2008-07-16 20:31:16 -07004669 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_DELAYEDACKLOST);
Pavel Emelyanov1ed83462008-07-16 20:29:51 -07004670 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004671
4672out_of_window:
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07004673 tcp_enter_quickack_mode(sk);
4674 inet_csk_schedule_ack(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004675drop:
4676 __kfree_skb(skb);
4677 return;
4678 }
4679
4680 /* Out of window. F.e. zero window probe. */
4681 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt + tcp_receive_window(tp)))
4682 goto out_of_window;
4683
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07004684 tcp_enter_quickack_mode(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004685
4686 if (before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
4687 /* Partial packet, seq < rcv_next < end_seq */
4688 SOCK_DEBUG(sk, "partial packet: rcv_next %X seq %X - %X\n",
4689 tp->rcv_nxt, TCP_SKB_CB(skb)->seq,
4690 TCP_SKB_CB(skb)->end_seq);
4691
Pavel Emelyanov1ed83462008-07-16 20:29:51 -07004692 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, tp->rcv_nxt);
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09004693
Linus Torvalds1da177e2005-04-16 15:20:36 -07004694 /* If window is closed, drop tail of packet. But after
4695 * remembering D-SACK for its head made in previous line.
4696 */
4697 if (!tcp_receive_window(tp))
4698 goto out_of_window;
4699 goto queue_and_out;
4700 }
4701
Eric Dumazete86b2912012-03-18 11:06:44 +00004702 tcp_data_queue_ofo(sk, skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004703}
4704
Ilpo Järvinen2cf46632008-08-23 05:11:41 -07004705static struct sk_buff *tcp_collapse_one(struct sock *sk, struct sk_buff *skb,
4706 struct sk_buff_head *list)
4707{
David S. Miller91521942009-05-28 21:35:47 -07004708 struct sk_buff *next = NULL;
4709
4710 if (!skb_queue_is_last(list, skb))
4711 next = skb_queue_next(list, skb);
Ilpo Järvinen2cf46632008-08-23 05:11:41 -07004712
4713 __skb_unlink(skb, list);
4714 __kfree_skb(skb);
4715 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPRCVCOLLAPSED);
4716
4717 return next;
4718}
4719
Linus Torvalds1da177e2005-04-16 15:20:36 -07004720/* Collapse contiguous sequence of skbs head..tail with
4721 * sequence numbers start..end.
David S. Miller91521942009-05-28 21:35:47 -07004722 *
4723 * If tail is NULL, this means until the end of the list.
4724 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07004725 * Segments with FIN/SYN are not collapsed (only because this
4726 * simplifies code)
4727 */
4728static void
David S. Miller8728b832005-08-09 19:25:21 -07004729tcp_collapse(struct sock *sk, struct sk_buff_head *list,
4730 struct sk_buff *head, struct sk_buff *tail,
4731 u32 start, u32 end)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004732{
David S. Miller91521942009-05-28 21:35:47 -07004733 struct sk_buff *skb, *n;
4734 bool end_of_skbs;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004735
Stephen Hemmingercaa20d9a2005-11-10 17:13:47 -08004736 /* First, check that queue is collapsible and find
Linus Torvalds1da177e2005-04-16 15:20:36 -07004737 * the point where collapsing can be useful. */
David S. Miller91521942009-05-28 21:35:47 -07004738 skb = head;
4739restart:
4740 end_of_skbs = true;
4741 skb_queue_walk_from_safe(list, skb, n) {
4742 if (skb == tail)
4743 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004744 /* No new bits? It is possible on ofo queue. */
4745 if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
Ilpo Järvinen2cf46632008-08-23 05:11:41 -07004746 skb = tcp_collapse_one(sk, skb, list);
David S. Miller91521942009-05-28 21:35:47 -07004747 if (!skb)
4748 break;
4749 goto restart;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004750 }
4751
4752 /* The first skb to collapse is:
4753 * - not SYN/FIN and
4754 * - bloated or contains data before "start" or
4755 * overlaps to the next one.
4756 */
Arnaldo Carvalho de Meloaa8223c2007-04-10 21:04:22 -07004757 if (!tcp_hdr(skb)->syn && !tcp_hdr(skb)->fin &&
Linus Torvalds1da177e2005-04-16 15:20:36 -07004758 (tcp_win_from_space(skb->truesize) > skb->len ||
David S. Miller91521942009-05-28 21:35:47 -07004759 before(TCP_SKB_CB(skb)->seq, start))) {
4760 end_of_skbs = false;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004761 break;
David S. Miller91521942009-05-28 21:35:47 -07004762 }
4763
4764 if (!skb_queue_is_last(list, skb)) {
4765 struct sk_buff *next = skb_queue_next(list, skb);
4766 if (next != tail &&
4767 TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(next)->seq) {
4768 end_of_skbs = false;
4769 break;
4770 }
4771 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07004772
4773 /* Decided to skip this, advance start seq. */
4774 start = TCP_SKB_CB(skb)->end_seq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004775 }
David S. Miller91521942009-05-28 21:35:47 -07004776 if (end_of_skbs || tcp_hdr(skb)->syn || tcp_hdr(skb)->fin)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004777 return;
4778
4779 while (before(start, end)) {
4780 struct sk_buff *nskb;
Chuck Leverc2636b42007-10-23 21:07:32 -07004781 unsigned int header = skb_headroom(skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004782 int copy = SKB_MAX_ORDER(header, 0);
4783
4784 /* Too big header? This can happen with IPv6. */
4785 if (copy < 0)
4786 return;
Ilpo Järvinen056834d2007-12-31 14:57:14 -08004787 if (end - start < copy)
4788 copy = end - start;
4789 nskb = alloc_skb(copy + header, GFP_ATOMIC);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004790 if (!nskb)
4791 return;
Arnaldo Carvalho de Meloc51957d2007-03-10 12:47:22 -03004792
Arnaldo Carvalho de Melo98e399f2007-03-19 15:33:04 -07004793 skb_set_mac_header(nskb, skb_mac_header(skb) - skb->head);
Arnaldo Carvalho de Melo9c702202007-04-25 18:04:18 -07004794 skb_set_network_header(nskb, (skb_network_header(skb) -
4795 skb->head));
4796 skb_set_transport_header(nskb, (skb_transport_header(skb) -
4797 skb->head));
Linus Torvalds1da177e2005-04-16 15:20:36 -07004798 skb_reserve(nskb, header);
4799 memcpy(nskb->head, skb->head, header);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004800 memcpy(nskb->cb, skb->cb, sizeof(skb->cb));
4801 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(nskb)->end_seq = start;
David S. Miller43f59c82008-09-21 21:28:51 -07004802 __skb_queue_before(list, skb, nskb);
Hideo Aoki3ab224b2007-12-31 00:11:19 -08004803 skb_set_owner_r(nskb, sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004804
4805 /* Copy data, releasing collapsed skbs. */
4806 while (copy > 0) {
4807 int offset = start - TCP_SKB_CB(skb)->seq;
4808 int size = TCP_SKB_CB(skb)->end_seq - start;
4809
Kris Katterjohn09a62662006-01-08 22:24:28 -08004810 BUG_ON(offset < 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004811 if (size > 0) {
4812 size = min(copy, size);
4813 if (skb_copy_bits(skb, offset, skb_put(nskb, size), size))
4814 BUG();
4815 TCP_SKB_CB(nskb)->end_seq += size;
4816 copy -= size;
4817 start += size;
4818 }
4819 if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
Ilpo Järvinen2cf46632008-08-23 05:11:41 -07004820 skb = tcp_collapse_one(sk, skb, list);
David S. Miller91521942009-05-28 21:35:47 -07004821 if (!skb ||
4822 skb == tail ||
Arnaldo Carvalho de Meloaa8223c2007-04-10 21:04:22 -07004823 tcp_hdr(skb)->syn ||
4824 tcp_hdr(skb)->fin)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004825 return;
4826 }
4827 }
4828 }
4829}
4830
4831/* Collapse ofo queue. Algorithm: select contiguous sequence of skbs
4832 * and tcp_collapse() them until all the queue is collapsed.
4833 */
4834static void tcp_collapse_ofo_queue(struct sock *sk)
4835{
4836 struct tcp_sock *tp = tcp_sk(sk);
4837 struct sk_buff *skb = skb_peek(&tp->out_of_order_queue);
4838 struct sk_buff *head;
4839 u32 start, end;
4840
4841 if (skb == NULL)
4842 return;
4843
4844 start = TCP_SKB_CB(skb)->seq;
4845 end = TCP_SKB_CB(skb)->end_seq;
4846 head = skb;
4847
4848 for (;;) {
David S. Miller91521942009-05-28 21:35:47 -07004849 struct sk_buff *next = NULL;
4850
4851 if (!skb_queue_is_last(&tp->out_of_order_queue, skb))
4852 next = skb_queue_next(&tp->out_of_order_queue, skb);
4853 skb = next;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004854
4855 /* Segment is terminated when we see gap or when
4856 * we are at the end of all the queue. */
David S. Miller91521942009-05-28 21:35:47 -07004857 if (!skb ||
Linus Torvalds1da177e2005-04-16 15:20:36 -07004858 after(TCP_SKB_CB(skb)->seq, end) ||
4859 before(TCP_SKB_CB(skb)->end_seq, start)) {
David S. Miller8728b832005-08-09 19:25:21 -07004860 tcp_collapse(sk, &tp->out_of_order_queue,
4861 head, skb, start, end);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004862 head = skb;
David S. Miller91521942009-05-28 21:35:47 -07004863 if (!skb)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004864 break;
4865 /* Start new segment */
4866 start = TCP_SKB_CB(skb)->seq;
4867 end = TCP_SKB_CB(skb)->end_seq;
4868 } else {
4869 if (before(TCP_SKB_CB(skb)->seq, start))
4870 start = TCP_SKB_CB(skb)->seq;
4871 if (after(TCP_SKB_CB(skb)->end_seq, end))
4872 end = TCP_SKB_CB(skb)->end_seq;
4873 }
4874 }
4875}
4876
Vitaliy Gusevb000cd32008-04-15 00:33:38 -07004877/*
4878 * Purge the out-of-order queue.
Vitaliy Gusev56f367b2008-04-15 20:26:34 -07004879 * Return true if queue was pruned.
Vitaliy Gusevb000cd32008-04-15 00:33:38 -07004880 */
Vitaliy Gusev56f367b2008-04-15 20:26:34 -07004881static int tcp_prune_ofo_queue(struct sock *sk)
Vitaliy Gusevb000cd32008-04-15 00:33:38 -07004882{
4883 struct tcp_sock *tp = tcp_sk(sk);
Vitaliy Gusev56f367b2008-04-15 20:26:34 -07004884 int res = 0;
Vitaliy Gusevb000cd32008-04-15 00:33:38 -07004885
4886 if (!skb_queue_empty(&tp->out_of_order_queue)) {
Pavel Emelyanovde0744a2008-07-16 20:31:16 -07004887 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_OFOPRUNED);
Vitaliy Gusevb000cd32008-04-15 00:33:38 -07004888 __skb_queue_purge(&tp->out_of_order_queue);
4889
4890 /* Reset SACK state. A conforming SACK implementation will
4891 * do the same at a timeout based retransmit. When a connection
4892 * is in a sad state like this, we care only about integrity
4893 * of the connection not performance.
4894 */
4895 if (tp->rx_opt.sack_ok)
4896 tcp_sack_reset(&tp->rx_opt);
4897 sk_mem_reclaim(sk);
Vitaliy Gusev56f367b2008-04-15 20:26:34 -07004898 res = 1;
Vitaliy Gusevb000cd32008-04-15 00:33:38 -07004899 }
Vitaliy Gusev56f367b2008-04-15 20:26:34 -07004900 return res;
Vitaliy Gusevb000cd32008-04-15 00:33:38 -07004901}
4902
Linus Torvalds1da177e2005-04-16 15:20:36 -07004903/* Reduce allocated memory if we can, trying to get
4904 * the socket within its memory limits again.
4905 *
4906 * Return less than zero if we should start dropping frames
4907 * until the socket owning process reads some of the data
4908 * to stabilize the situation.
4909 */
4910static int tcp_prune_queue(struct sock *sk)
4911{
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09004912 struct tcp_sock *tp = tcp_sk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004913
4914 SOCK_DEBUG(sk, "prune_queue: c=%x\n", tp->copied_seq);
4915
Pavel Emelyanovde0744a2008-07-16 20:31:16 -07004916 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_PRUNECALLED);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004917
4918 if (atomic_read(&sk->sk_rmem_alloc) >= sk->sk_rcvbuf)
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07004919 tcp_clamp_window(sk);
Glauber Costa180d8cd2011-12-11 21:47:02 +00004920 else if (sk_under_memory_pressure(sk))
Linus Torvalds1da177e2005-04-16 15:20:36 -07004921 tp->rcv_ssthresh = min(tp->rcv_ssthresh, 4U * tp->advmss);
4922
4923 tcp_collapse_ofo_queue(sk);
David S. Miller91521942009-05-28 21:35:47 -07004924 if (!skb_queue_empty(&sk->sk_receive_queue))
4925 tcp_collapse(sk, &sk->sk_receive_queue,
4926 skb_peek(&sk->sk_receive_queue),
4927 NULL,
4928 tp->copied_seq, tp->rcv_nxt);
Hideo Aoki3ab224b2007-12-31 00:11:19 -08004929 sk_mem_reclaim(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004930
4931 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
4932 return 0;
4933
4934 /* Collapsing did not help, destructive actions follow.
4935 * This must not ever occur. */
4936
Vitaliy Gusevb000cd32008-04-15 00:33:38 -07004937 tcp_prune_ofo_queue(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004938
4939 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
4940 return 0;
4941
4942 /* If we are really being abused, tell the caller to silently
4943 * drop receive data on the floor. It will get retransmitted
4944 * and hopefully then we'll have sufficient space.
4945 */
Pavel Emelyanovde0744a2008-07-16 20:31:16 -07004946 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_RCVPRUNED);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004947
4948 /* Massive buffer overcommit. */
4949 tp->pred_flags = 0;
4950 return -1;
4951}
4952
Linus Torvalds1da177e2005-04-16 15:20:36 -07004953/* RFC2861, slow part. Adjust cwnd, after it was not full during one rto.
4954 * As additional protections, we do not touch cwnd in retransmission phases,
4955 * and if application hit its sndbuf limit recently.
4956 */
4957void tcp_cwnd_application_limited(struct sock *sk)
4958{
4959 struct tcp_sock *tp = tcp_sk(sk);
4960
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03004961 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Open &&
Linus Torvalds1da177e2005-04-16 15:20:36 -07004962 sk->sk_socket && !test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) {
4963 /* Limited by application or receiver window. */
Ilpo Järvinend254bcd2006-08-04 16:57:42 -07004964 u32 init_win = tcp_init_cwnd(tp, __sk_dst_get(sk));
4965 u32 win_used = max(tp->snd_cwnd_used, init_win);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004966 if (win_used < tp->snd_cwnd) {
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03004967 tp->snd_ssthresh = tcp_current_ssthresh(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004968 tp->snd_cwnd = (tp->snd_cwnd + win_used) >> 1;
4969 }
4970 tp->snd_cwnd_used = 0;
4971 }
4972 tp->snd_cwnd_stamp = tcp_time_stamp;
4973}
4974
Eric Dumazetcf533ea2011-10-21 05:22:42 -04004975static int tcp_should_expand_sndbuf(const struct sock *sk)
David S. Miller0d9901d2005-07-05 15:21:10 -07004976{
Eric Dumazetcf533ea2011-10-21 05:22:42 -04004977 const struct tcp_sock *tp = tcp_sk(sk);
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07004978
David S. Miller0d9901d2005-07-05 15:21:10 -07004979 /* If the user specified a specific send buffer setting, do
4980 * not modify it.
4981 */
4982 if (sk->sk_userlocks & SOCK_SNDBUF_LOCK)
4983 return 0;
4984
4985 /* If we are under global TCP memory pressure, do not expand. */
Glauber Costa180d8cd2011-12-11 21:47:02 +00004986 if (sk_under_memory_pressure(sk))
David S. Miller0d9901d2005-07-05 15:21:10 -07004987 return 0;
4988
4989 /* If we are under soft global TCP memory pressure, do not expand. */
Glauber Costa180d8cd2011-12-11 21:47:02 +00004990 if (sk_memory_allocated(sk) >= sk_prot_mem_limits(sk, 0))
David S. Miller0d9901d2005-07-05 15:21:10 -07004991 return 0;
4992
4993 /* If we filled the congestion window, do not expand. */
4994 if (tp->packets_out >= tp->snd_cwnd)
4995 return 0;
4996
4997 return 1;
4998}
Linus Torvalds1da177e2005-04-16 15:20:36 -07004999
5000/* When incoming ACK allowed to free some skb from write_queue,
5001 * we remember this event in flag SOCK_QUEUE_SHRUNK and wake up socket
5002 * on the exit from tcp input handler.
5003 *
5004 * PROBLEM: sndbuf expansion does not work well with largesend.
5005 */
5006static void tcp_new_space(struct sock *sk)
5007{
5008 struct tcp_sock *tp = tcp_sk(sk);
5009
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07005010 if (tcp_should_expand_sndbuf(sk)) {
Eric Dumazet87fb4b72011-10-13 07:28:54 +00005011 int sndmem = SKB_TRUESIZE(max_t(u32,
5012 tp->rx_opt.mss_clamp,
5013 tp->mss_cache) +
5014 MAX_TCP_HEADER);
Ilpo Järvinen4a7e5602008-10-07 14:43:31 -07005015 int demanded = max_t(unsigned int, tp->snd_cwnd,
Ilpo Järvinen056834d2007-12-31 14:57:14 -08005016 tp->reordering + 1);
5017 sndmem *= 2 * demanded;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005018 if (sndmem > sk->sk_sndbuf)
5019 sk->sk_sndbuf = min(sndmem, sysctl_tcp_wmem[2]);
5020 tp->snd_cwnd_stamp = tcp_time_stamp;
5021 }
5022
5023 sk->sk_write_space(sk);
5024}
5025
Stephen Hemminger40efc6f2006-01-03 16:03:49 -08005026static void tcp_check_space(struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005027{
5028 if (sock_flag(sk, SOCK_QUEUE_SHRUNK)) {
5029 sock_reset_flag(sk, SOCK_QUEUE_SHRUNK);
5030 if (sk->sk_socket &&
5031 test_bit(SOCK_NOSPACE, &sk->sk_socket->flags))
5032 tcp_new_space(sk);
5033 }
5034}
5035
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07005036static inline void tcp_data_snd_check(struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005037{
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07005038 tcp_push_pending_frames(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005039 tcp_check_space(sk);
5040}
5041
5042/*
5043 * Check if sending an ack is needed.
5044 */
5045static void __tcp_ack_snd_check(struct sock *sk, int ofo_possible)
5046{
5047 struct tcp_sock *tp = tcp_sk(sk);
5048
5049 /* More than one full frame received... */
Joe Perches9d4fb272009-11-23 10:41:23 -08005050 if (((tp->rcv_nxt - tp->rcv_wup) > inet_csk(sk)->icsk_ack.rcv_mss &&
Linus Torvalds1da177e2005-04-16 15:20:36 -07005051 /* ... and right edge of window advances far enough.
5052 * (tcp_recvmsg() will send ACK otherwise). Or...
5053 */
Joe Perches9d4fb272009-11-23 10:41:23 -08005054 __tcp_select_window(sk) >= tp->rcv_wnd) ||
Linus Torvalds1da177e2005-04-16 15:20:36 -07005055 /* We ACK each frame or... */
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07005056 tcp_in_quickack_mode(sk) ||
Linus Torvalds1da177e2005-04-16 15:20:36 -07005057 /* We have out of order data. */
Ilpo Järvinen056834d2007-12-31 14:57:14 -08005058 (ofo_possible && skb_peek(&tp->out_of_order_queue))) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07005059 /* Then ack it now */
5060 tcp_send_ack(sk);
5061 } else {
5062 /* Else, send delayed ack. */
5063 tcp_send_delayed_ack(sk);
5064 }
5065}
5066
Stephen Hemminger40efc6f2006-01-03 16:03:49 -08005067static inline void tcp_ack_snd_check(struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005068{
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07005069 if (!inet_csk_ack_scheduled(sk)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07005070 /* We sent a data segment already. */
5071 return;
5072 }
5073 __tcp_ack_snd_check(sk, 1);
5074}
5075
5076/*
5077 * This routine is only called when we have urgent data
Stephen Hemmingercaa20d9a2005-11-10 17:13:47 -08005078 * signaled. Its the 'slow' part of tcp_urg. It could be
Linus Torvalds1da177e2005-04-16 15:20:36 -07005079 * moved inline now as tcp_urg is only called from one
5080 * place. We handle URGent data wrong. We have to - as
5081 * BSD still doesn't use the correction from RFC961.
5082 * For 1003.1g we should support a new option TCP_STDURG to permit
5083 * either form (or just set the sysctl tcp_stdurg).
5084 */
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09005085
Eric Dumazetcf533ea2011-10-21 05:22:42 -04005086static void tcp_check_urg(struct sock *sk, const struct tcphdr *th)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005087{
5088 struct tcp_sock *tp = tcp_sk(sk);
5089 u32 ptr = ntohs(th->urg_ptr);
5090
5091 if (ptr && !sysctl_tcp_stdurg)
5092 ptr--;
5093 ptr += ntohl(th->seq);
5094
5095 /* Ignore urgent data that we've already seen and read. */
5096 if (after(tp->copied_seq, ptr))
5097 return;
5098
5099 /* Do not replay urg ptr.
5100 *
5101 * NOTE: interesting situation not covered by specs.
5102 * Misbehaving sender may send urg ptr, pointing to segment,
5103 * which we already have in ofo queue. We are not able to fetch
5104 * such data and will stay in TCP_URG_NOTYET until will be eaten
5105 * by recvmsg(). Seems, we are not obliged to handle such wicked
5106 * situations. But it is worth to think about possibility of some
5107 * DoSes using some hypothetical application level deadlock.
5108 */
5109 if (before(ptr, tp->rcv_nxt))
5110 return;
5111
5112 /* Do we already have a newer (or duplicate) urgent pointer? */
5113 if (tp->urg_data && !after(ptr, tp->urg_seq))
5114 return;
5115
5116 /* Tell the world about our new urgent pointer. */
5117 sk_send_sigurg(sk);
5118
5119 /* We may be adding urgent data when the last byte read was
5120 * urgent. To do this requires some care. We cannot just ignore
5121 * tp->copied_seq since we would read the last urgent byte again
5122 * as data, nor can we alter copied_seq until this data arrives
Stephen Hemmingercaa20d9a2005-11-10 17:13:47 -08005123 * or we break the semantics of SIOCATMARK (and thus sockatmark())
Linus Torvalds1da177e2005-04-16 15:20:36 -07005124 *
5125 * NOTE. Double Dutch. Rendering to plain English: author of comment
5126 * above did something sort of send("A", MSG_OOB); send("B", MSG_OOB);
5127 * and expect that both A and B disappear from stream. This is _wrong_.
5128 * Though this happens in BSD with high probability, this is occasional.
5129 * Any application relying on this is buggy. Note also, that fix "works"
5130 * only in this artificial test. Insert some normal data between A and B and we will
5131 * decline of BSD again. Verdict: it is better to remove to trap
5132 * buggy users.
5133 */
5134 if (tp->urg_seq == tp->copied_seq && tp->urg_data &&
Ilpo Järvinen056834d2007-12-31 14:57:14 -08005135 !sock_flag(sk, SOCK_URGINLINE) && tp->copied_seq != tp->rcv_nxt) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07005136 struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
5137 tp->copied_seq++;
5138 if (skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq)) {
David S. Miller8728b832005-08-09 19:25:21 -07005139 __skb_unlink(skb, &sk->sk_receive_queue);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005140 __kfree_skb(skb);
5141 }
5142 }
5143
Ilpo Järvinen056834d2007-12-31 14:57:14 -08005144 tp->urg_data = TCP_URG_NOTYET;
5145 tp->urg_seq = ptr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005146
5147 /* Disable header prediction. */
5148 tp->pred_flags = 0;
5149}
5150
5151/* This is the 'fast' part of urgent handling. */
Eric Dumazetcf533ea2011-10-21 05:22:42 -04005152static void tcp_urg(struct sock *sk, struct sk_buff *skb, const struct tcphdr *th)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005153{
5154 struct tcp_sock *tp = tcp_sk(sk);
5155
5156 /* Check if we get a new urgent pointer - normally not. */
5157 if (th->urg)
Ilpo Järvinen056834d2007-12-31 14:57:14 -08005158 tcp_check_urg(sk, th);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005159
5160 /* Do we wait for any urgent data? - normally not... */
5161 if (tp->urg_data == TCP_URG_NOTYET) {
5162 u32 ptr = tp->urg_seq - ntohl(th->seq) + (th->doff * 4) -
5163 th->syn;
5164
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09005165 /* Is the urgent pointer pointing into this packet? */
Linus Torvalds1da177e2005-04-16 15:20:36 -07005166 if (ptr < skb->len) {
5167 u8 tmp;
5168 if (skb_copy_bits(skb, ptr, &tmp, 1))
5169 BUG();
5170 tp->urg_data = TCP_URG_VALID | tmp;
5171 if (!sock_flag(sk, SOCK_DEAD))
5172 sk->sk_data_ready(sk, 0);
5173 }
5174 }
5175}
5176
5177static int tcp_copy_to_iovec(struct sock *sk, struct sk_buff *skb, int hlen)
5178{
5179 struct tcp_sock *tp = tcp_sk(sk);
5180 int chunk = skb->len - hlen;
5181 int err;
5182
5183 local_bh_enable();
Herbert Xu60476372007-04-09 11:59:39 -07005184 if (skb_csum_unnecessary(skb))
Linus Torvalds1da177e2005-04-16 15:20:36 -07005185 err = skb_copy_datagram_iovec(skb, hlen, tp->ucopy.iov, chunk);
5186 else
5187 err = skb_copy_and_csum_datagram_iovec(skb, hlen,
5188 tp->ucopy.iov);
5189
5190 if (!err) {
5191 tp->ucopy.len -= chunk;
5192 tp->copied_seq += chunk;
5193 tcp_rcv_space_adjust(sk);
5194 }
5195
5196 local_bh_disable();
5197 return err;
5198}
5199
Ilpo Järvinen056834d2007-12-31 14:57:14 -08005200static __sum16 __tcp_checksum_complete_user(struct sock *sk,
5201 struct sk_buff *skb)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005202{
Al Virob51655b2006-11-14 21:40:42 -08005203 __sum16 result;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005204
5205 if (sock_owned_by_user(sk)) {
5206 local_bh_enable();
5207 result = __tcp_checksum_complete(skb);
5208 local_bh_disable();
5209 } else {
5210 result = __tcp_checksum_complete(skb);
5211 }
5212 return result;
5213}
5214
Ilpo Järvinen056834d2007-12-31 14:57:14 -08005215static inline int tcp_checksum_complete_user(struct sock *sk,
5216 struct sk_buff *skb)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005217{
Herbert Xu60476372007-04-09 11:59:39 -07005218 return !skb_csum_unnecessary(skb) &&
Ilpo Järvinen056834d2007-12-31 14:57:14 -08005219 __tcp_checksum_complete_user(sk, skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005220}
5221
Chris Leech1a2449a2006-05-23 18:05:53 -07005222#ifdef CONFIG_NET_DMA
Ilpo Järvinen056834d2007-12-31 14:57:14 -08005223static int tcp_dma_try_early_copy(struct sock *sk, struct sk_buff *skb,
5224 int hlen)
Chris Leech1a2449a2006-05-23 18:05:53 -07005225{
5226 struct tcp_sock *tp = tcp_sk(sk);
5227 int chunk = skb->len - hlen;
5228 int dma_cookie;
5229 int copied_early = 0;
5230
5231 if (tp->ucopy.wakeup)
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09005232 return 0;
Chris Leech1a2449a2006-05-23 18:05:53 -07005233
5234 if (!tp->ucopy.dma_chan && tp->ucopy.pinned_list)
Dave Jianga2bd1142012-04-04 16:10:46 -07005235 tp->ucopy.dma_chan = net_dma_find_channel();
Chris Leech1a2449a2006-05-23 18:05:53 -07005236
Herbert Xu60476372007-04-09 11:59:39 -07005237 if (tp->ucopy.dma_chan && skb_csum_unnecessary(skb)) {
Chris Leech1a2449a2006-05-23 18:05:53 -07005238
5239 dma_cookie = dma_skb_copy_datagram_iovec(tp->ucopy.dma_chan,
Ilpo Järvinen056834d2007-12-31 14:57:14 -08005240 skb, hlen,
5241 tp->ucopy.iov, chunk,
5242 tp->ucopy.pinned_list);
Chris Leech1a2449a2006-05-23 18:05:53 -07005243
5244 if (dma_cookie < 0)
5245 goto out;
5246
5247 tp->ucopy.dma_cookie = dma_cookie;
5248 copied_early = 1;
5249
5250 tp->ucopy.len -= chunk;
5251 tp->copied_seq += chunk;
5252 tcp_rcv_space_adjust(sk);
5253
5254 if ((tp->ucopy.len == 0) ||
Arnaldo Carvalho de Meloaa8223c2007-04-10 21:04:22 -07005255 (tcp_flag_word(tcp_hdr(skb)) & TCP_FLAG_PSH) ||
Chris Leech1a2449a2006-05-23 18:05:53 -07005256 (atomic_read(&sk->sk_rmem_alloc) > (sk->sk_rcvbuf >> 1))) {
5257 tp->ucopy.wakeup = 1;
5258 sk->sk_data_ready(sk, 0);
5259 }
5260 } else if (chunk > 0) {
5261 tp->ucopy.wakeup = 1;
5262 sk->sk_data_ready(sk, 0);
5263 }
5264out:
5265 return copied_early;
5266}
5267#endif /* CONFIG_NET_DMA */
5268
Ilpo Järvinencbe2d122008-08-23 05:10:12 -07005269/* Does PAWS and seqno based validation of an incoming segment, flags will
5270 * play significant role here.
5271 */
5272static int tcp_validate_incoming(struct sock *sk, struct sk_buff *skb,
Eric Dumazetcf533ea2011-10-21 05:22:42 -04005273 const struct tcphdr *th, int syn_inerr)
Ilpo Järvinencbe2d122008-08-23 05:10:12 -07005274{
Eric Dumazetcf533ea2011-10-21 05:22:42 -04005275 const u8 *hash_location;
Ilpo Järvinencbe2d122008-08-23 05:10:12 -07005276 struct tcp_sock *tp = tcp_sk(sk);
5277
5278 /* RFC1323: H1. Apply PAWS check first. */
William Allen Simpson4957faa2009-12-02 18:25:27 +00005279 if (tcp_fast_parse_options(skb, th, tp, &hash_location) &&
5280 tp->rx_opt.saw_tstamp &&
Ilpo Järvinencbe2d122008-08-23 05:10:12 -07005281 tcp_paws_discard(sk, skb)) {
5282 if (!th->rst) {
5283 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_PAWSESTABREJECTED);
5284 tcp_send_dupack(sk, skb);
5285 goto discard;
5286 }
5287 /* Reset is accepted even if it did not pass PAWS. */
5288 }
5289
5290 /* Step 1: check sequence number */
5291 if (!tcp_sequence(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)) {
5292 /* RFC793, page 37: "In all states except SYN-SENT, all reset
5293 * (RST) segments are validated by checking their SEQ-fields."
5294 * And page 69: "If an incoming segment is not acceptable,
5295 * an acknowledgment should be sent in reply (unless the RST
5296 * bit is set, if so drop the segment and return)".
5297 */
5298 if (!th->rst)
5299 tcp_send_dupack(sk, skb);
5300 goto discard;
5301 }
5302
5303 /* Step 2: check RST bit */
5304 if (th->rst) {
5305 tcp_reset(sk);
5306 goto discard;
5307 }
5308
5309 /* ts_recent update must be made after we are sure that the packet
5310 * is in window.
5311 */
5312 tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq);
5313
5314 /* step 3: check security and precedence [ignored] */
5315
5316 /* step 4: Check for a SYN in window. */
5317 if (th->syn && !before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
5318 if (syn_inerr)
5319 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_INERRS);
5320 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPABORTONSYN);
5321 tcp_reset(sk);
5322 return -1;
5323 }
5324
5325 return 1;
5326
5327discard:
5328 __kfree_skb(skb);
5329 return 0;
5330}
5331
Linus Torvalds1da177e2005-04-16 15:20:36 -07005332/*
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09005333 * TCP receive function for the ESTABLISHED state.
Linus Torvalds1da177e2005-04-16 15:20:36 -07005334 *
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09005335 * It is split into a fast path and a slow path. The fast path is
Linus Torvalds1da177e2005-04-16 15:20:36 -07005336 * disabled when:
5337 * - A zero window was announced from us - zero window probing
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09005338 * is only handled properly in the slow path.
Linus Torvalds1da177e2005-04-16 15:20:36 -07005339 * - Out of order segments arrived.
5340 * - Urgent data is expected.
5341 * - There is no buffer space left
5342 * - Unexpected TCP flags/window values/header lengths are received
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09005343 * (detected by checking the TCP header against pred_flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005344 * - Data is sent in both directions. Fast path only supports pure senders
5345 * or pure receivers (this means either the sequence number or the ack
5346 * value must stay constant)
5347 * - Unexpected TCP option.
5348 *
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09005349 * When these conditions are not satisfied it drops into a standard
Linus Torvalds1da177e2005-04-16 15:20:36 -07005350 * receive procedure patterned after RFC793 to handle all cases.
5351 * The first three cases are guaranteed by proper pred_flags setting,
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09005352 * the rest is checked inline. Fast processing is turned on in
Linus Torvalds1da177e2005-04-16 15:20:36 -07005353 * tcp_data_queue when everything is OK.
5354 */
5355int tcp_rcv_established(struct sock *sk, struct sk_buff *skb,
Eric Dumazetcf533ea2011-10-21 05:22:42 -04005356 const struct tcphdr *th, unsigned int len)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005357{
5358 struct tcp_sock *tp = tcp_sk(sk);
Ilpo Järvinencbe2d122008-08-23 05:10:12 -07005359 int res;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005360
5361 /*
5362 * Header prediction.
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09005363 * The code loosely follows the one in the famous
Linus Torvalds1da177e2005-04-16 15:20:36 -07005364 * "30 instruction TCP receive" Van Jacobson mail.
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09005365 *
5366 * Van's trick is to deposit buffers into socket queue
Linus Torvalds1da177e2005-04-16 15:20:36 -07005367 * on a device interrupt, to call tcp_recv function
5368 * on the receive process context and checksum and copy
5369 * the buffer to user space. smart...
5370 *
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09005371 * Our current scheme is not silly either but we take the
Linus Torvalds1da177e2005-04-16 15:20:36 -07005372 * extra cost of the net_bh soft interrupt processing...
5373 * We do checksum and copy also but from device to kernel.
5374 */
5375
5376 tp->rx_opt.saw_tstamp = 0;
5377
5378 /* pred_flags is 0xS?10 << 16 + snd_wnd
Stephen Hemmingercaa20d9a2005-11-10 17:13:47 -08005379 * if header_prediction is to be made
Linus Torvalds1da177e2005-04-16 15:20:36 -07005380 * 'S' will always be tp->tcp_header_len >> 2
5381 * '?' will be 0 for the fast path, otherwise pred_flags is 0 to
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09005382 * turn it off (when there are holes in the receive
Linus Torvalds1da177e2005-04-16 15:20:36 -07005383 * space for instance)
5384 * PSH flag is ignored.
5385 */
5386
5387 if ((tcp_flag_word(th) & TCP_HP_BITS) == tp->pred_flags &&
John Dykstra96e0bf42009-03-22 21:49:57 -07005388 TCP_SKB_CB(skb)->seq == tp->rcv_nxt &&
5389 !after(TCP_SKB_CB(skb)->ack_seq, tp->snd_nxt)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07005390 int tcp_header_len = tp->tcp_header_len;
5391
5392 /* Timestamp header prediction: tcp_header_len
5393 * is automatically equal to th->doff*4 due to pred_flags
5394 * match.
5395 */
5396
5397 /* Check timestamp */
5398 if (tcp_header_len == sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07005399 /* No? Slow path! */
Ilpo Järvinena4356b22008-08-23 05:12:29 -07005400 if (!tcp_parse_aligned_timestamp(tp, th))
Linus Torvalds1da177e2005-04-16 15:20:36 -07005401 goto slow_path;
5402
Linus Torvalds1da177e2005-04-16 15:20:36 -07005403 /* If PAWS failed, check it more carefully in slow path */
5404 if ((s32)(tp->rx_opt.rcv_tsval - tp->rx_opt.ts_recent) < 0)
5405 goto slow_path;
5406
5407 /* DO NOT update ts_recent here, if checksum fails
5408 * and timestamp was corrupted part, it will result
5409 * in a hung connection since we will drop all
5410 * future packets due to the PAWS test.
5411 */
5412 }
5413
5414 if (len <= tcp_header_len) {
5415 /* Bulk data transfer: sender */
5416 if (len == tcp_header_len) {
5417 /* Predicted packet is in window by definition.
5418 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5419 * Hence, check seq<=rcv_wup reduces to:
5420 */
5421 if (tcp_header_len ==
5422 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
5423 tp->rcv_nxt == tp->rcv_wup)
5424 tcp_store_ts_recent(tp);
5425
Linus Torvalds1da177e2005-04-16 15:20:36 -07005426 /* We know that such packets are checksummed
5427 * on entry.
5428 */
5429 tcp_ack(sk, skb, 0);
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09005430 __kfree_skb(skb);
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07005431 tcp_data_snd_check(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005432 return 0;
5433 } else { /* Header too small */
Pavel Emelyanov63231bd2008-07-16 20:22:25 -07005434 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_INERRS);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005435 goto discard;
5436 }
5437 } else {
5438 int eaten = 0;
Chris Leech1a2449a2006-05-23 18:05:53 -07005439 int copied_early = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005440
Chris Leech1a2449a2006-05-23 18:05:53 -07005441 if (tp->copied_seq == tp->rcv_nxt &&
5442 len - tcp_header_len <= tp->ucopy.len) {
5443#ifdef CONFIG_NET_DMA
5444 if (tcp_dma_try_early_copy(sk, skb, tcp_header_len)) {
5445 copied_early = 1;
5446 eaten = 1;
5447 }
5448#endif
Ilpo Järvinen056834d2007-12-31 14:57:14 -08005449 if (tp->ucopy.task == current &&
5450 sock_owned_by_user(sk) && !copied_early) {
Chris Leech1a2449a2006-05-23 18:05:53 -07005451 __set_current_state(TASK_RUNNING);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005452
Chris Leech1a2449a2006-05-23 18:05:53 -07005453 if (!tcp_copy_to_iovec(sk, skb, tcp_header_len))
5454 eaten = 1;
5455 }
5456 if (eaten) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07005457 /* Predicted packet is in window by definition.
5458 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5459 * Hence, check seq<=rcv_wup reduces to:
5460 */
5461 if (tcp_header_len ==
5462 (sizeof(struct tcphdr) +
5463 TCPOLEN_TSTAMP_ALIGNED) &&
5464 tp->rcv_nxt == tp->rcv_wup)
5465 tcp_store_ts_recent(tp);
5466
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07005467 tcp_rcv_rtt_measure_ts(sk, skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005468
5469 __skb_pull(skb, tcp_header_len);
5470 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
Pavel Emelyanovde0744a2008-07-16 20:31:16 -07005471 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPHPHITSTOUSER);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005472 }
Chris Leech1a2449a2006-05-23 18:05:53 -07005473 if (copied_early)
5474 tcp_cleanup_rbuf(sk, skb->len);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005475 }
5476 if (!eaten) {
5477 if (tcp_checksum_complete_user(sk, skb))
5478 goto csum_error;
5479
5480 /* Predicted packet is in window by definition.
5481 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5482 * Hence, check seq<=rcv_wup reduces to:
5483 */
5484 if (tcp_header_len ==
5485 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
5486 tp->rcv_nxt == tp->rcv_wup)
5487 tcp_store_ts_recent(tp);
5488
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07005489 tcp_rcv_rtt_measure_ts(sk, skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005490
5491 if ((int)skb->truesize > sk->sk_forward_alloc)
5492 goto step5;
5493
Pavel Emelyanovde0744a2008-07-16 20:31:16 -07005494 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPHPHITS);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005495
5496 /* Bulk data transfer: receiver */
Ilpo Järvinen056834d2007-12-31 14:57:14 -08005497 __skb_pull(skb, tcp_header_len);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005498 __skb_queue_tail(&sk->sk_receive_queue, skb);
Hideo Aoki3ab224b2007-12-31 00:11:19 -08005499 skb_set_owner_r(skb, sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005500 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
5501 }
5502
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07005503 tcp_event_data_recv(sk, skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005504
5505 if (TCP_SKB_CB(skb)->ack_seq != tp->snd_una) {
5506 /* Well, only one small jumplet in fast path... */
5507 tcp_ack(sk, skb, FLAG_DATA);
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07005508 tcp_data_snd_check(sk);
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07005509 if (!inet_csk_ack_scheduled(sk))
Linus Torvalds1da177e2005-04-16 15:20:36 -07005510 goto no_ack;
5511 }
5512
Ali Saidi53240c22008-10-07 15:31:19 -07005513 if (!copied_early || tp->rcv_nxt != tp->rcv_wup)
5514 __tcp_ack_snd_check(sk, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005515no_ack:
Chris Leech1a2449a2006-05-23 18:05:53 -07005516#ifdef CONFIG_NET_DMA
5517 if (copied_early)
5518 __skb_queue_tail(&sk->sk_async_wait_queue, skb);
5519 else
5520#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07005521 if (eaten)
5522 __kfree_skb(skb);
5523 else
5524 sk->sk_data_ready(sk, 0);
5525 return 0;
5526 }
5527 }
5528
5529slow_path:
Ilpo Järvinen056834d2007-12-31 14:57:14 -08005530 if (len < (th->doff << 2) || tcp_checksum_complete_user(sk, skb))
Linus Torvalds1da177e2005-04-16 15:20:36 -07005531 goto csum_error;
5532
5533 /*
Linus Torvalds1da177e2005-04-16 15:20:36 -07005534 * Standard slow path.
5535 */
5536
Ilpo Järvinencbe2d122008-08-23 05:10:12 -07005537 res = tcp_validate_incoming(sk, skb, th, 1);
5538 if (res <= 0)
5539 return -res;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005540
5541step5:
John Dykstra96e0bf42009-03-22 21:49:57 -07005542 if (th->ack && tcp_ack(sk, skb, FLAG_SLOWPATH) < 0)
5543 goto discard;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005544
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07005545 tcp_rcv_rtt_measure_ts(sk, skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005546
5547 /* Process urgent data. */
5548 tcp_urg(sk, skb, th);
5549
5550 /* step 7: process the segment text */
5551 tcp_data_queue(sk, skb);
5552
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07005553 tcp_data_snd_check(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005554 tcp_ack_snd_check(sk);
5555 return 0;
5556
5557csum_error:
Pavel Emelyanov63231bd2008-07-16 20:22:25 -07005558 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_INERRS);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005559
5560discard:
5561 __kfree_skb(skb);
5562 return 0;
5563}
Eric Dumazet4bc2f182010-07-09 21:22:10 +00005564EXPORT_SYMBOL(tcp_rcv_established);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005565
5566static int tcp_rcv_synsent_state_process(struct sock *sk, struct sk_buff *skb,
Eric Dumazetcf533ea2011-10-21 05:22:42 -04005567 const struct tcphdr *th, unsigned int len)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005568{
Eric Dumazetcf533ea2011-10-21 05:22:42 -04005569 const u8 *hash_location;
Arnaldo Carvalho de Melod83d8462005-12-13 23:26:10 -08005570 struct inet_connection_sock *icsk = inet_csk(sk);
William Allen Simpson4957faa2009-12-02 18:25:27 +00005571 struct tcp_sock *tp = tcp_sk(sk);
William Allen Simpson4957faa2009-12-02 18:25:27 +00005572 struct tcp_cookie_values *cvp = tp->cookie_values;
5573 int saved_clamp = tp->rx_opt.mss_clamp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005574
David S. Millerbb5b7c12009-12-15 20:56:42 -08005575 tcp_parse_options(skb, &tp->rx_opt, &hash_location, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005576
5577 if (th->ack) {
5578 /* rfc793:
5579 * "If the state is SYN-SENT then
5580 * first check the ACK bit
5581 * If the ACK bit is set
5582 * If SEG.ACK =< ISS, or SEG.ACK > SND.NXT, send
5583 * a reset (unless the RST bit is set, if so drop
5584 * the segment and return)"
5585 *
5586 * We do not send data with SYN, so that RFC-correct
5587 * test reduces to:
5588 */
5589 if (TCP_SKB_CB(skb)->ack_seq != tp->snd_nxt)
5590 goto reset_and_undo;
5591
5592 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
5593 !between(tp->rx_opt.rcv_tsecr, tp->retrans_stamp,
5594 tcp_time_stamp)) {
Pavel Emelyanovde0744a2008-07-16 20:31:16 -07005595 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_PAWSACTIVEREJECTED);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005596 goto reset_and_undo;
5597 }
5598
5599 /* Now ACK is acceptable.
5600 *
5601 * "If the RST bit is set
5602 * If the ACK was acceptable then signal the user "error:
5603 * connection reset", drop the segment, enter CLOSED state,
5604 * delete TCB, and return."
5605 */
5606
5607 if (th->rst) {
5608 tcp_reset(sk);
5609 goto discard;
5610 }
5611
5612 /* rfc793:
5613 * "fifth, if neither of the SYN or RST bits is set then
5614 * drop the segment and return."
5615 *
5616 * See note below!
5617 * --ANK(990513)
5618 */
5619 if (!th->syn)
5620 goto discard_and_undo;
5621
5622 /* rfc793:
5623 * "If the SYN bit is on ...
5624 * are acceptable then ...
5625 * (our SYN has been ACKed), change the connection
5626 * state to ESTABLISHED..."
5627 */
5628
5629 TCP_ECN_rcv_synack(tp, th);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005630
5631 tp->snd_wl1 = TCP_SKB_CB(skb)->seq;
5632 tcp_ack(sk, skb, FLAG_SLOWPATH);
5633
5634 /* Ok.. it's good. Set up sequence numbers and
5635 * move to established.
5636 */
5637 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
5638 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
5639
5640 /* RFC1323: The window in SYN & SYN/ACK segments is
5641 * never scaled.
5642 */
5643 tp->snd_wnd = ntohs(th->window);
Hantzis Fotisee7537b2009-03-02 22:42:02 -08005644 tcp_init_wl(tp, TCP_SKB_CB(skb)->seq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005645
5646 if (!tp->rx_opt.wscale_ok) {
5647 tp->rx_opt.snd_wscale = tp->rx_opt.rcv_wscale = 0;
5648 tp->window_clamp = min(tp->window_clamp, 65535U);
5649 }
5650
5651 if (tp->rx_opt.saw_tstamp) {
5652 tp->rx_opt.tstamp_ok = 1;
5653 tp->tcp_header_len =
5654 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
5655 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
5656 tcp_store_ts_recent(tp);
5657 } else {
5658 tp->tcp_header_len = sizeof(struct tcphdr);
5659 }
5660
Ilpo Järvinene60402d2007-08-09 15:14:46 +03005661 if (tcp_is_sack(tp) && sysctl_tcp_fack)
5662 tcp_enable_fack(tp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005663
John Heffner5d424d52006-03-20 17:53:41 -08005664 tcp_mtup_init(sk);
Arnaldo Carvalho de Melod83d8462005-12-13 23:26:10 -08005665 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005666 tcp_initialize_rcv_mss(sk);
5667
5668 /* Remember, tcp_poll() does not lock socket!
5669 * Change state from SYN-SENT only after copied_seq
5670 * is initialized. */
5671 tp->copied_seq = tp->rcv_nxt;
William Allen Simpson4957faa2009-12-02 18:25:27 +00005672
5673 if (cvp != NULL &&
5674 cvp->cookie_pair_size > 0 &&
5675 tp->rx_opt.cookie_plus > 0) {
5676 int cookie_size = tp->rx_opt.cookie_plus
5677 - TCPOLEN_COOKIE_BASE;
5678 int cookie_pair_size = cookie_size
5679 + cvp->cookie_desired;
5680
5681 /* A cookie extension option was sent and returned.
5682 * Note that each incoming SYNACK replaces the
5683 * Responder cookie. The initial exchange is most
5684 * fragile, as protection against spoofing relies
5685 * entirely upon the sequence and timestamp (above).
5686 * This replacement strategy allows the correct pair to
5687 * pass through, while any others will be filtered via
5688 * Responder verification later.
5689 */
5690 if (sizeof(cvp->cookie_pair) >= cookie_pair_size) {
5691 memcpy(&cvp->cookie_pair[cvp->cookie_desired],
5692 hash_location, cookie_size);
5693 cvp->cookie_pair_size = cookie_pair_size;
5694 }
5695 }
5696
Ralf Baechlee16aa202006-12-07 00:11:33 -08005697 smp_mb();
Linus Torvalds1da177e2005-04-16 15:20:36 -07005698 tcp_set_state(sk, TCP_ESTABLISHED);
5699
Venkat Yekkirala6b877692006-11-08 17:04:09 -06005700 security_inet_conn_established(sk, skb);
5701
Linus Torvalds1da177e2005-04-16 15:20:36 -07005702 /* Make sure socket is routed, for correct metrics. */
Arnaldo Carvalho de Melo8292a172005-12-13 23:15:52 -08005703 icsk->icsk_af_ops->rebuild_header(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005704
5705 tcp_init_metrics(sk);
5706
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03005707 tcp_init_congestion_control(sk);
Stephen Hemminger317a76f2005-06-23 12:19:55 -07005708
Linus Torvalds1da177e2005-04-16 15:20:36 -07005709 /* Prevent spurious tcp_cwnd_restart() on first data
5710 * packet.
5711 */
5712 tp->lsndtime = tcp_time_stamp;
5713
5714 tcp_init_buffer_space(sk);
5715
5716 if (sock_flag(sk, SOCK_KEEPOPEN))
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07005717 inet_csk_reset_keepalive_timer(sk, keepalive_time_when(tp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07005718
5719 if (!tp->rx_opt.snd_wscale)
5720 __tcp_fast_path_on(tp, tp->snd_wnd);
5721 else
5722 tp->pred_flags = 0;
5723
5724 if (!sock_flag(sk, SOCK_DEAD)) {
5725 sk->sk_state_change(sk);
Pavel Emelyanov8d8ad9d2007-11-26 20:10:50 +08005726 sk_wake_async(sk, SOCK_WAKE_IO, POLL_OUT);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005727 }
5728
Arnaldo Carvalho de Melo295f7322005-08-09 20:11:56 -07005729 if (sk->sk_write_pending ||
5730 icsk->icsk_accept_queue.rskq_defer_accept ||
5731 icsk->icsk_ack.pingpong) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07005732 /* Save one ACK. Data will be ready after
5733 * several ticks, if write_pending is set.
5734 *
5735 * It may be deleted, but with this feature tcpdumps
5736 * look so _wonderfully_ clever, that I was not able
5737 * to stand against the temptation 8) --ANK
5738 */
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07005739 inet_csk_schedule_ack(sk);
Arnaldo Carvalho de Melo295f7322005-08-09 20:11:56 -07005740 icsk->icsk_ack.lrcvtime = tcp_time_stamp;
5741 icsk->icsk_ack.ato = TCP_ATO_MIN;
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07005742 tcp_incr_quickack(sk);
5743 tcp_enter_quickack_mode(sk);
Arnaldo Carvalho de Melo3f421ba2005-08-09 20:11:08 -07005744 inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
5745 TCP_DELACK_MAX, TCP_RTO_MAX);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005746
5747discard:
5748 __kfree_skb(skb);
5749 return 0;
5750 } else {
5751 tcp_send_ack(sk);
5752 }
5753 return -1;
5754 }
5755
5756 /* No ACK in the segment */
5757
5758 if (th->rst) {
5759 /* rfc793:
5760 * "If the RST bit is set
5761 *
5762 * Otherwise (no ACK) drop the segment and return."
5763 */
5764
5765 goto discard_and_undo;
5766 }
5767
5768 /* PAWS check. */
Ilpo Järvinen056834d2007-12-31 14:57:14 -08005769 if (tp->rx_opt.ts_recent_stamp && tp->rx_opt.saw_tstamp &&
Ilpo Järvinenc887e6d2009-03-14 14:23:03 +00005770 tcp_paws_reject(&tp->rx_opt, 0))
Linus Torvalds1da177e2005-04-16 15:20:36 -07005771 goto discard_and_undo;
5772
5773 if (th->syn) {
5774 /* We see SYN without ACK. It is attempt of
5775 * simultaneous connect with crossed SYNs.
5776 * Particularly, it can be connect to self.
5777 */
5778 tcp_set_state(sk, TCP_SYN_RECV);
5779
5780 if (tp->rx_opt.saw_tstamp) {
5781 tp->rx_opt.tstamp_ok = 1;
5782 tcp_store_ts_recent(tp);
5783 tp->tcp_header_len =
5784 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
5785 } else {
5786 tp->tcp_header_len = sizeof(struct tcphdr);
5787 }
5788
5789 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
5790 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
5791
5792 /* RFC1323: The window in SYN & SYN/ACK segments is
5793 * never scaled.
5794 */
5795 tp->snd_wnd = ntohs(th->window);
5796 tp->snd_wl1 = TCP_SKB_CB(skb)->seq;
5797 tp->max_window = tp->snd_wnd;
5798
5799 TCP_ECN_rcv_syn(tp, th);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005800
John Heffner5d424d52006-03-20 17:53:41 -08005801 tcp_mtup_init(sk);
Arnaldo Carvalho de Melod83d8462005-12-13 23:26:10 -08005802 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005803 tcp_initialize_rcv_mss(sk);
5804
Linus Torvalds1da177e2005-04-16 15:20:36 -07005805 tcp_send_synack(sk);
5806#if 0
5807 /* Note, we could accept data and URG from this segment.
5808 * There are no obstacles to make this.
5809 *
5810 * However, if we ignore data in ACKless segments sometimes,
5811 * we have no reasons to accept it sometimes.
5812 * Also, seems the code doing it in step6 of tcp_rcv_state_process
5813 * is not flawless. So, discard packet for sanity.
5814 * Uncomment this return to process the data.
5815 */
5816 return -1;
5817#else
5818 goto discard;
5819#endif
5820 }
5821 /* "fifth, if neither of the SYN or RST bits is set then
5822 * drop the segment and return."
5823 */
5824
5825discard_and_undo:
5826 tcp_clear_options(&tp->rx_opt);
5827 tp->rx_opt.mss_clamp = saved_clamp;
5828 goto discard;
5829
5830reset_and_undo:
5831 tcp_clear_options(&tp->rx_opt);
5832 tp->rx_opt.mss_clamp = saved_clamp;
5833 return 1;
5834}
5835
Linus Torvalds1da177e2005-04-16 15:20:36 -07005836/*
5837 * This function implements the receiving procedure of RFC 793 for
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09005838 * all states except ESTABLISHED and TIME_WAIT.
Linus Torvalds1da177e2005-04-16 15:20:36 -07005839 * It's called from both tcp_v4_rcv and tcp_v6_rcv and should be
5840 * address independent.
5841 */
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09005842
Linus Torvalds1da177e2005-04-16 15:20:36 -07005843int tcp_rcv_state_process(struct sock *sk, struct sk_buff *skb,
Eric Dumazetcf533ea2011-10-21 05:22:42 -04005844 const struct tcphdr *th, unsigned int len)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005845{
5846 struct tcp_sock *tp = tcp_sk(sk);
Arnaldo Carvalho de Melo8292a172005-12-13 23:15:52 -08005847 struct inet_connection_sock *icsk = inet_csk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005848 int queued = 0;
Ilpo Järvinencbe2d122008-08-23 05:10:12 -07005849 int res;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005850
5851 tp->rx_opt.saw_tstamp = 0;
5852
5853 switch (sk->sk_state) {
5854 case TCP_CLOSE:
5855 goto discard;
5856
5857 case TCP_LISTEN:
Stephen Hemminger2de979b2007-03-08 20:45:19 -08005858 if (th->ack)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005859 return 1;
5860
Stephen Hemminger2de979b2007-03-08 20:45:19 -08005861 if (th->rst)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005862 goto discard;
5863
Stephen Hemminger2de979b2007-03-08 20:45:19 -08005864 if (th->syn) {
Eric Dumazetfdf5af02011-12-02 23:41:42 +00005865 if (th->fin)
5866 goto discard;
Arnaldo Carvalho de Melo8292a172005-12-13 23:15:52 -08005867 if (icsk->icsk_af_ops->conn_request(sk, skb) < 0)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005868 return 1;
5869
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09005870 /* Now we have several options: In theory there is
5871 * nothing else in the frame. KA9Q has an option to
Linus Torvalds1da177e2005-04-16 15:20:36 -07005872 * send data with the syn, BSD accepts data with the
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09005873 * syn up to the [to be] advertised window and
5874 * Solaris 2.1 gives you a protocol error. For now
5875 * we just ignore it, that fits the spec precisely
Linus Torvalds1da177e2005-04-16 15:20:36 -07005876 * and avoids incompatibilities. It would be nice in
5877 * future to drop through and process the data.
5878 *
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09005879 * Now that TTCP is starting to be used we ought to
Linus Torvalds1da177e2005-04-16 15:20:36 -07005880 * queue this data.
5881 * But, this leaves one open to an easy denial of
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09005882 * service attack, and SYN cookies can't defend
Linus Torvalds1da177e2005-04-16 15:20:36 -07005883 * against this problem. So, we drop the data
Masayuki Nakagawafb7e2392007-01-23 20:15:06 -08005884 * in the interest of security over speed unless
5885 * it's still in use.
Linus Torvalds1da177e2005-04-16 15:20:36 -07005886 */
Masayuki Nakagawafb7e2392007-01-23 20:15:06 -08005887 kfree_skb(skb);
5888 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005889 }
5890 goto discard;
5891
5892 case TCP_SYN_SENT:
Linus Torvalds1da177e2005-04-16 15:20:36 -07005893 queued = tcp_rcv_synsent_state_process(sk, skb, th, len);
5894 if (queued >= 0)
5895 return queued;
5896
5897 /* Do step6 onward by hand. */
5898 tcp_urg(sk, skb, th);
5899 __kfree_skb(skb);
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07005900 tcp_data_snd_check(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005901 return 0;
5902 }
5903
Ilpo Järvinencbe2d122008-08-23 05:10:12 -07005904 res = tcp_validate_incoming(sk, skb, th, 0);
5905 if (res <= 0)
5906 return -res;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005907
5908 /* step 5: check the ACK field */
5909 if (th->ack) {
John Dykstra96e0bf42009-03-22 21:49:57 -07005910 int acceptable = tcp_ack(sk, skb, FLAG_SLOWPATH) > 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005911
Stephen Hemminger2de979b2007-03-08 20:45:19 -08005912 switch (sk->sk_state) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07005913 case TCP_SYN_RECV:
5914 if (acceptable) {
5915 tp->copied_seq = tp->rcv_nxt;
Ralf Baechlee16aa202006-12-07 00:11:33 -08005916 smp_mb();
Linus Torvalds1da177e2005-04-16 15:20:36 -07005917 tcp_set_state(sk, TCP_ESTABLISHED);
5918 sk->sk_state_change(sk);
5919
5920 /* Note, that this wakeup is only for marginal
5921 * crossed SYN case. Passively open sockets
5922 * are not waked up, because sk->sk_sleep ==
5923 * NULL and sk->sk_socket == NULL.
5924 */
Pavel Emelyanov8d8ad9d2007-11-26 20:10:50 +08005925 if (sk->sk_socket)
5926 sk_wake_async(sk,
Ilpo Järvinen056834d2007-12-31 14:57:14 -08005927 SOCK_WAKE_IO, POLL_OUT);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005928
5929 tp->snd_una = TCP_SKB_CB(skb)->ack_seq;
5930 tp->snd_wnd = ntohs(th->window) <<
5931 tp->rx_opt.snd_wscale;
Hantzis Fotisee7537b2009-03-02 22:42:02 -08005932 tcp_init_wl(tp, TCP_SKB_CB(skb)->seq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005933
Linus Torvalds1da177e2005-04-16 15:20:36 -07005934 if (tp->rx_opt.tstamp_ok)
5935 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
5936
5937 /* Make sure socket is routed, for
5938 * correct metrics.
5939 */
Arnaldo Carvalho de Melo8292a172005-12-13 23:15:52 -08005940 icsk->icsk_af_ops->rebuild_header(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005941
5942 tcp_init_metrics(sk);
5943
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03005944 tcp_init_congestion_control(sk);
Stephen Hemminger317a76f2005-06-23 12:19:55 -07005945
Linus Torvalds1da177e2005-04-16 15:20:36 -07005946 /* Prevent spurious tcp_cwnd_restart() on
5947 * first data packet.
5948 */
5949 tp->lsndtime = tcp_time_stamp;
5950
John Heffner5d424d52006-03-20 17:53:41 -08005951 tcp_mtup_init(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005952 tcp_initialize_rcv_mss(sk);
5953 tcp_init_buffer_space(sk);
5954 tcp_fast_path_on(tp);
5955 } else {
5956 return 1;
5957 }
5958 break;
5959
5960 case TCP_FIN_WAIT1:
5961 if (tp->snd_una == tp->write_seq) {
5962 tcp_set_state(sk, TCP_FIN_WAIT2);
5963 sk->sk_shutdown |= SEND_SHUTDOWN;
Eric Dumazetb6c67122010-04-08 23:03:29 +00005964 dst_confirm(__sk_dst_get(sk));
Linus Torvalds1da177e2005-04-16 15:20:36 -07005965
5966 if (!sock_flag(sk, SOCK_DEAD))
5967 /* Wake up lingering close() */
5968 sk->sk_state_change(sk);
5969 else {
5970 int tmo;
5971
5972 if (tp->linger2 < 0 ||
5973 (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
5974 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt))) {
5975 tcp_done(sk);
Pavel Emelyanovde0744a2008-07-16 20:31:16 -07005976 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005977 return 1;
5978 }
5979
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07005980 tmo = tcp_fin_time(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005981 if (tmo > TCP_TIMEWAIT_LEN) {
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07005982 inet_csk_reset_keepalive_timer(sk, tmo - TCP_TIMEWAIT_LEN);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005983 } else if (th->fin || sock_owned_by_user(sk)) {
5984 /* Bad case. We could lose such FIN otherwise.
5985 * It is not a big problem, but it looks confusing
5986 * and not so rare event. We still can lose it now,
5987 * if it spins in bh_lock_sock(), but it is really
5988 * marginal case.
5989 */
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07005990 inet_csk_reset_keepalive_timer(sk, tmo);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005991 } else {
5992 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
5993 goto discard;
5994 }
5995 }
5996 }
5997 break;
5998
5999 case TCP_CLOSING:
6000 if (tp->snd_una == tp->write_seq) {
6001 tcp_time_wait(sk, TCP_TIME_WAIT, 0);
6002 goto discard;
6003 }
6004 break;
6005
6006 case TCP_LAST_ACK:
6007 if (tp->snd_una == tp->write_seq) {
6008 tcp_update_metrics(sk);
6009 tcp_done(sk);
6010 goto discard;
6011 }
6012 break;
6013 }
6014 } else
6015 goto discard;
6016
6017 /* step 6: check the URG bit */
6018 tcp_urg(sk, skb, th);
6019
6020 /* step 7: process the segment text */
6021 switch (sk->sk_state) {
6022 case TCP_CLOSE_WAIT:
6023 case TCP_CLOSING:
6024 case TCP_LAST_ACK:
6025 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
6026 break;
6027 case TCP_FIN_WAIT1:
6028 case TCP_FIN_WAIT2:
6029 /* RFC 793 says to queue data in these states,
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09006030 * RFC 1122 says we MUST send a reset.
Linus Torvalds1da177e2005-04-16 15:20:36 -07006031 * BSD 4.4 also does reset.
6032 */
6033 if (sk->sk_shutdown & RCV_SHUTDOWN) {
6034 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
6035 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt)) {
Pavel Emelyanovde0744a2008-07-16 20:31:16 -07006036 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006037 tcp_reset(sk);
6038 return 1;
6039 }
6040 }
6041 /* Fall through */
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09006042 case TCP_ESTABLISHED:
Linus Torvalds1da177e2005-04-16 15:20:36 -07006043 tcp_data_queue(sk, skb);
6044 queued = 1;
6045 break;
6046 }
6047
6048 /* tcp_data could move socket to TIME-WAIT */
6049 if (sk->sk_state != TCP_CLOSE) {
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07006050 tcp_data_snd_check(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006051 tcp_ack_snd_check(sk);
6052 }
6053
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09006054 if (!queued) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07006055discard:
6056 __kfree_skb(skb);
6057 }
6058 return 0;
6059}
Linus Torvalds1da177e2005-04-16 15:20:36 -07006060EXPORT_SYMBOL(tcp_rcv_state_process);